program(1.0)
[buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.7.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})]
{
    func main<ios17>(tensor<int32, [1, ?]> codes) [FlexibleShapeInformation = tuple<tuple<tensor<string, []>, dict<tensor<string, []>, tensor<int32, [?]>>>, tuple<tensor<string, []>, dict<tensor<string, []>, list<tensor<int32, [2]>, ?>>>>((("DefaultShapes", {{"codes", [1, 500]}}), ("RangeDims", {{"codes", [[1, 1], [2, 2000]]}})))] {
            tensor<fp32, [1024]> final_layer_norm_bias = const()[name = tensor<string, []>("final_layer_norm_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64)))];
            tensor<fp32, [1024]> final_layer_norm_weight = const()[name = tensor<string, []>("final_layer_norm_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4224)))];
            tensor<int32, []> input_3_batch_dims_0 = const()[name = tensor<string, []>("input_3_batch_dims_0"), val = tensor<int32, []>(0)];
            tensor<bool, []> input_3_validate_indices_0 = const()[name = tensor<string, []>("input_3_validate_indices_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [65536, 8]> fsq_table_weight_to_fp16 = const()[name = tensor<string, []>("fsq_table_weight_to_fp16"), val = tensor<fp16, [65536, 8]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8384)))];
            tensor<int32, []> greater_equal_0_y_0 = const()[name = tensor<string, []>("greater_equal_0_y_0"), val = tensor<int32, []>(0)];
            tensor<bool, [1, ?]> greater_equal_0 = greater_equal(x = codes, y = greater_equal_0_y_0)[name = tensor<string, []>("greater_equal_0")];
            tensor<int32, []> slice_by_index_48 = const()[name = tensor<string, []>("slice_by_index_48"), val = tensor<int32, []>(65536)];
            tensor<int32, [1, ?]> add_16 = add(x = codes, y = slice_by_index_48)[name = tensor<string, []>("add_16")];
            tensor<int32, [1, ?]> select_0 = select(a = codes, b = add_16, cond = greater_equal_0)[name = tensor<string, []>("select_0")];
            tensor<int32, []> input_3_cast_fp16_axis_0 = const()[name = tensor<string, []>("input_3_cast_fp16_axis_0"), val = tensor<int32, []>(0)];
            tensor<fp16, [1, ?, 8]> input_3_cast_fp16 = gather(axis = input_3_cast_fp16_axis_0, batch_dims = input_3_batch_dims_0, indices = select_0, validate_indices = input_3_validate_indices_0, x = fsq_table_weight_to_fp16)[name = tensor<string, []>("input_3_cast_fp16")];
            tensor<fp16, [2048, 8]> fsq_project_out_weight_to_fp16 = const()[name = tensor<string, []>("fsq_project_out_weight_to_fp16"), val = tensor<fp16, [2048, 8]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1057024)))];
            tensor<fp16, [2048]> fsq_project_out_bias_to_fp16 = const()[name = tensor<string, []>("fsq_project_out_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1089856)))];
            tensor<fp16, [1, ?, 2048]> linear_0_cast_fp16 = linear(bias = fsq_project_out_bias_to_fp16, weight = fsq_project_out_weight_to_fp16, x = input_3_cast_fp16)[name = tensor<string, []>("linear_0_cast_fp16")];
            tensor<fp16, [1024, 2048]> fc_post_a_weight_to_fp16 = const()[name = tensor<string, []>("fc_post_a_weight_to_fp16"), val = tensor<fp16, [1024, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1094016)))];
            tensor<fp16, [1024]> fc_post_a_bias_to_fp16 = const()[name = tensor<string, []>("fc_post_a_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5288384)))];
            tensor<fp16, [1, ?, 1024]> linear_1_cast_fp16 = linear(bias = fc_post_a_bias_to_fp16, weight = fc_post_a_weight_to_fp16, x = linear_0_cast_fp16)[name = tensor<string, []>("linear_1_cast_fp16")];
            tensor<int32, [3]> input_7_perm_0 = const()[name = tensor<string, []>("input_7_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> input_9_pad_type_0 = const()[name = tensor<string, []>("input_9_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [2]> input_9_pad_0 = const()[name = tensor<string, []>("input_9_pad_0"), val = tensor<int32, [2]>([3, 3])];
            tensor<int32, [1]> input_9_strides_0 = const()[name = tensor<string, []>("input_9_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> input_9_dilations_0 = const()[name = tensor<string, []>("input_9_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> input_9_groups_0 = const()[name = tensor<string, []>("input_9_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 7]> embed_weight_to_fp16 = const()[name = tensor<string, []>("embed_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 7]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5290496)))];
            tensor<fp16, [1024]> embed_bias_to_fp16 = const()[name = tensor<string, []>("embed_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19970624)))];
            tensor<fp16, [1, 1024, ?]> input_7_cast_fp16 = transpose(perm = input_7_perm_0, x = linear_1_cast_fp16)[name = tensor<string, []>("transpose_52")];
            tensor<fp16, [1, 1024, ?]> input_9_cast_fp16 = conv(bias = embed_bias_to_fp16, dilations = input_9_dilations_0, groups = input_9_groups_0, pad = input_9_pad_0, pad_type = input_9_pad_type_0, strides = input_9_strides_0, weight = embed_weight_to_fp16, x = input_7_cast_fp16)[name = tensor<string, []>("input_9_cast_fp16")];
            tensor<int32, [3]> shape_0_cast_fp16 = shape(x = input_9_cast_fp16)[name = tensor<string, []>("shape_0_cast_fp16")];
            tensor<int32, [4]> concat_0x = const()[name = tensor<string, []>("concat_0x"), val = tensor<int32, [4]>([1, 32, 32, -1])];
            tensor<fp16, [1, 32, 32, ?]> reshape_0_cast_fp16 = reshape(shape = concat_0x, x = input_9_cast_fp16)[name = tensor<string, []>("reshape_0_cast_fp16")];
            tensor<string, []> reshape_0_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("reshape_0_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<int32, [2]> reduce_mean_0_axes_0 = const()[name = tensor<string, []>("reduce_mean_0_axes_0"), val = tensor<int32, [2]>([2, 3])];
            tensor<bool, []> reduce_mean_0_keep_dims_0 = const()[name = tensor<string, []>("reduce_mean_0_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 32, 32, ?]> reshape_0_cast_fp16_to_fp32 = cast(dtype = reshape_0_cast_fp16_to_fp32_dtype_0, x = reshape_0_cast_fp16)[name = tensor<string, []>("cast_234")];
            tensor<fp32, [1, 32, 1, 1]> reduce_mean_0 = reduce_mean(axes = reduce_mean_0_axes_0, keep_dims = reduce_mean_0_keep_dims_0, x = reshape_0_cast_fp16_to_fp32)[name = tensor<string, []>("reduce_mean_0")];
            tensor<string, []> reduce_mean_0_to_fp16_dtype_0 = const()[name = tensor<string, []>("reduce_mean_0_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 32, 1, 1]> reduce_mean_0_to_fp16 = cast(dtype = reduce_mean_0_to_fp16_dtype_0, x = reduce_mean_0)[name = tensor<string, []>("cast_233")];
            tensor<fp16, [1, 32, 32, ?]> sub_0_cast_fp16 = sub(x = reshape_0_cast_fp16, y = reduce_mean_0_to_fp16)[name = tensor<string, []>("sub_0_cast_fp16")];
            tensor<fp16, [1, 32, 32, ?]> square_0_cast_fp16 = square(x = sub_0_cast_fp16)[name = tensor<string, []>("square_0_cast_fp16")];
            tensor<string, []> square_0_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("square_0_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<int32, [2]> reduce_mean_2_axes_0 = const()[name = tensor<string, []>("reduce_mean_2_axes_0"), val = tensor<int32, [2]>([2, 3])];
            tensor<bool, []> reduce_mean_2_keep_dims_0 = const()[name = tensor<string, []>("reduce_mean_2_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 32, 32, ?]> square_0_cast_fp16_to_fp32 = cast(dtype = square_0_cast_fp16_to_fp32_dtype_0, x = square_0_cast_fp16)[name = tensor<string, []>("cast_232")];
            tensor<fp32, [1, 32, 1, 1]> reduce_mean_2 = reduce_mean(axes = reduce_mean_2_axes_0, keep_dims = reduce_mean_2_keep_dims_0, x = square_0_cast_fp16_to_fp32)[name = tensor<string, []>("reduce_mean_2")];
            tensor<string, []> reduce_mean_2_to_fp16_dtype_0 = const()[name = tensor<string, []>("reduce_mean_2_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> add_0_y_0_to_fp16 = const()[name = tensor<string, []>("add_0_y_0_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 32, 1, 1]> reduce_mean_2_to_fp16 = cast(dtype = reduce_mean_2_to_fp16_dtype_0, x = reduce_mean_2)[name = tensor<string, []>("cast_231")];
            tensor<fp16, [1, 32, 1, 1]> add_0_cast_fp16 = add(x = reduce_mean_2_to_fp16, y = add_0_y_0_to_fp16)[name = tensor<string, []>("add_0_cast_fp16")];
            tensor<fp16, [1, 32, 1, 1]> sqrt_0_cast_fp16 = sqrt(x = add_0_cast_fp16)[name = tensor<string, []>("sqrt_0_cast_fp16")];
            tensor<fp16, [1, 32, 32, ?]> real_div_0_cast_fp16 = real_div(x = sub_0_cast_fp16, y = sqrt_0_cast_fp16)[name = tensor<string, []>("real_div_0_cast_fp16")];
            tensor<fp16, [1, 1024, ?]> reshape_1_cast_fp16 = reshape(shape = shape_0_cast_fp16, x = real_div_0_cast_fp16)[name = tensor<string, []>("reshape_1_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> reshape_2_to_fp16 = const()[name = tensor<string, []>("reshape_2_to_fp16"), val = tensor<fp16, [1, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19972736)))];
            tensor<fp16, [1, 1024, ?]> mul_0_cast_fp16 = mul(x = reshape_1_cast_fp16, y = reshape_2_to_fp16)[name = tensor<string, []>("mul_0_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> reshape_3_to_fp16 = const()[name = tensor<string, []>("reshape_3_to_fp16"), val = tensor<fp16, [1, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19974848)))];
            tensor<fp16, [1, 1024, ?]> add_1_cast_fp16 = add(x = mul_0_cast_fp16, y = reshape_3_to_fp16)[name = tensor<string, []>("add_1_cast_fp16")];
            tensor<fp16, [1, 1024, ?]> input_11_cast_fp16 = silu(x = add_1_cast_fp16)[name = tensor<string, []>("input_11_cast_fp16")];
            tensor<string, []> input_13_pad_type_0 = const()[name = tensor<string, []>("input_13_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [2]> input_13_pad_0 = const()[name = tensor<string, []>("input_13_pad_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [1]> input_13_strides_0 = const()[name = tensor<string, []>("input_13_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> input_13_dilations_0 = const()[name = tensor<string, []>("input_13_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> input_13_groups_0 = const()[name = tensor<string, []>("input_13_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 3]> prior_net_0_conv1_weight_to_fp16 = const()[name = tensor<string, []>("prior_net_0_conv1_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 3]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19976960)))];
            tensor<fp16, [1024]> prior_net_0_conv1_bias_to_fp16 = const()[name = tensor<string, []>("prior_net_0_conv1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(26268480)))];
            tensor<fp16, [1, 1024, ?]> input_13_cast_fp16 = conv(bias = prior_net_0_conv1_bias_to_fp16, dilations = input_13_dilations_0, groups = input_13_groups_0, pad = input_13_pad_0, pad_type = input_13_pad_type_0, strides = input_13_strides_0, weight = prior_net_0_conv1_weight_to_fp16, x = input_11_cast_fp16)[name = tensor<string, []>("input_13_cast_fp16")];
            tensor<int32, [3]> shape_1_cast_fp16 = shape(x = input_13_cast_fp16)[name = tensor<string, []>("shape_1_cast_fp16")];
            tensor<int32, [4]> concat_1x = const()[name = tensor<string, []>("concat_1x"), val = tensor<int32, [4]>([1, 32, 32, -1])];
            tensor<fp16, [1, 32, 32, ?]> reshape_4_cast_fp16 = reshape(shape = concat_1x, x = input_13_cast_fp16)[name = tensor<string, []>("reshape_4_cast_fp16")];
            tensor<string, []> reshape_4_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("reshape_4_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<int32, [2]> reduce_mean_3_axes_0 = const()[name = tensor<string, []>("reduce_mean_3_axes_0"), val = tensor<int32, [2]>([2, 3])];
            tensor<bool, []> reduce_mean_3_keep_dims_0 = const()[name = tensor<string, []>("reduce_mean_3_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 32, 32, ?]> reshape_4_cast_fp16_to_fp32 = cast(dtype = reshape_4_cast_fp16_to_fp32_dtype_0, x = reshape_4_cast_fp16)[name = tensor<string, []>("cast_230")];
            tensor<fp32, [1, 32, 1, 1]> reduce_mean_3 = reduce_mean(axes = reduce_mean_3_axes_0, keep_dims = reduce_mean_3_keep_dims_0, x = reshape_4_cast_fp16_to_fp32)[name = tensor<string, []>("reduce_mean_3")];
            tensor<string, []> reduce_mean_3_to_fp16_dtype_0 = const()[name = tensor<string, []>("reduce_mean_3_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 32, 1, 1]> reduce_mean_3_to_fp16 = cast(dtype = reduce_mean_3_to_fp16_dtype_0, x = reduce_mean_3)[name = tensor<string, []>("cast_229")];
            tensor<fp16, [1, 32, 32, ?]> sub_2_cast_fp16 = sub(x = reshape_4_cast_fp16, y = reduce_mean_3_to_fp16)[name = tensor<string, []>("sub_2_cast_fp16")];
            tensor<fp16, [1, 32, 32, ?]> square_1_cast_fp16 = square(x = sub_2_cast_fp16)[name = tensor<string, []>("square_1_cast_fp16")];
            tensor<string, []> square_1_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("square_1_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<int32, [2]> reduce_mean_5_axes_0 = const()[name = tensor<string, []>("reduce_mean_5_axes_0"), val = tensor<int32, [2]>([2, 3])];
            tensor<bool, []> reduce_mean_5_keep_dims_0 = const()[name = tensor<string, []>("reduce_mean_5_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 32, 32, ?]> square_1_cast_fp16_to_fp32 = cast(dtype = square_1_cast_fp16_to_fp32_dtype_0, x = square_1_cast_fp16)[name = tensor<string, []>("cast_228")];
            tensor<fp32, [1, 32, 1, 1]> reduce_mean_5 = reduce_mean(axes = reduce_mean_5_axes_0, keep_dims = reduce_mean_5_keep_dims_0, x = square_1_cast_fp16_to_fp32)[name = tensor<string, []>("reduce_mean_5")];
            tensor<string, []> reduce_mean_5_to_fp16_dtype_0 = const()[name = tensor<string, []>("reduce_mean_5_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> add_2_y_0_to_fp16 = const()[name = tensor<string, []>("add_2_y_0_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 32, 1, 1]> reduce_mean_5_to_fp16 = cast(dtype = reduce_mean_5_to_fp16_dtype_0, x = reduce_mean_5)[name = tensor<string, []>("cast_227")];
            tensor<fp16, [1, 32, 1, 1]> add_2_cast_fp16 = add(x = reduce_mean_5_to_fp16, y = add_2_y_0_to_fp16)[name = tensor<string, []>("add_2_cast_fp16")];
            tensor<fp16, [1, 32, 1, 1]> sqrt_1_cast_fp16 = sqrt(x = add_2_cast_fp16)[name = tensor<string, []>("sqrt_1_cast_fp16")];
            tensor<fp16, [1, 32, 32, ?]> real_div_1_cast_fp16 = real_div(x = sub_2_cast_fp16, y = sqrt_1_cast_fp16)[name = tensor<string, []>("real_div_1_cast_fp16")];
            tensor<fp16, [1, 1024, ?]> reshape_5_cast_fp16 = reshape(shape = shape_1_cast_fp16, x = real_div_1_cast_fp16)[name = tensor<string, []>("reshape_5_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> reshape_6_to_fp16 = const()[name = tensor<string, []>("reshape_6_to_fp16"), val = tensor<fp16, [1, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(26270592)))];
            tensor<fp16, [1, 1024, ?]> mul_1_cast_fp16 = mul(x = reshape_5_cast_fp16, y = reshape_6_to_fp16)[name = tensor<string, []>("mul_1_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> reshape_7_to_fp16 = const()[name = tensor<string, []>("reshape_7_to_fp16"), val = tensor<fp16, [1, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(26272704)))];
            tensor<fp16, [1, 1024, ?]> add_3_cast_fp16 = add(x = mul_1_cast_fp16, y = reshape_7_to_fp16)[name = tensor<string, []>("add_3_cast_fp16")];
            tensor<fp16, [1, 1024, ?]> input_15_cast_fp16 = silu(x = add_3_cast_fp16)[name = tensor<string, []>("input_15_cast_fp16")];
            tensor<string, []> h_1_pad_type_0 = const()[name = tensor<string, []>("h_1_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [2]> h_1_pad_0 = const()[name = tensor<string, []>("h_1_pad_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [1]> h_1_strides_0 = const()[name = tensor<string, []>("h_1_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> h_1_dilations_0 = const()[name = tensor<string, []>("h_1_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> h_1_groups_0 = const()[name = tensor<string, []>("h_1_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 3]> prior_net_0_conv2_weight_to_fp16 = const()[name = tensor<string, []>("prior_net_0_conv2_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 3]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(26274816)))];
            tensor<fp16, [1024]> prior_net_0_conv2_bias_to_fp16 = const()[name = tensor<string, []>("prior_net_0_conv2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(32566336)))];
            tensor<fp16, [1, 1024, ?]> h_1_cast_fp16 = conv(bias = prior_net_0_conv2_bias_to_fp16, dilations = h_1_dilations_0, groups = h_1_groups_0, pad = h_1_pad_0, pad_type = h_1_pad_type_0, strides = h_1_strides_0, weight = prior_net_0_conv2_weight_to_fp16, x = input_15_cast_fp16)[name = tensor<string, []>("h_1_cast_fp16")];
            tensor<fp16, [1, 1024, ?]> input_19_cast_fp16 = add(x = input_9_cast_fp16, y = h_1_cast_fp16)[name = tensor<string, []>("input_19_cast_fp16")];
            tensor<int32, [3]> shape_2_cast_fp16 = shape(x = input_19_cast_fp16)[name = tensor<string, []>("shape_2_cast_fp16")];
            tensor<int32, [4]> concat_2x = const()[name = tensor<string, []>("concat_2x"), val = tensor<int32, [4]>([1, 32, 32, -1])];
            tensor<fp16, [1, 32, 32, ?]> reshape_8_cast_fp16 = reshape(shape = concat_2x, x = input_19_cast_fp16)[name = tensor<string, []>("reshape_8_cast_fp16")];
            tensor<string, []> reshape_8_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("reshape_8_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<int32, [2]> reduce_mean_6_axes_0 = const()[name = tensor<string, []>("reduce_mean_6_axes_0"), val = tensor<int32, [2]>([2, 3])];
            tensor<bool, []> reduce_mean_6_keep_dims_0 = const()[name = tensor<string, []>("reduce_mean_6_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 32, 32, ?]> reshape_8_cast_fp16_to_fp32 = cast(dtype = reshape_8_cast_fp16_to_fp32_dtype_0, x = reshape_8_cast_fp16)[name = tensor<string, []>("cast_226")];
            tensor<fp32, [1, 32, 1, 1]> reduce_mean_6 = reduce_mean(axes = reduce_mean_6_axes_0, keep_dims = reduce_mean_6_keep_dims_0, x = reshape_8_cast_fp16_to_fp32)[name = tensor<string, []>("reduce_mean_6")];
            tensor<string, []> reduce_mean_6_to_fp16_dtype_0 = const()[name = tensor<string, []>("reduce_mean_6_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 32, 1, 1]> reduce_mean_6_to_fp16 = cast(dtype = reduce_mean_6_to_fp16_dtype_0, x = reduce_mean_6)[name = tensor<string, []>("cast_225")];
            tensor<fp16, [1, 32, 32, ?]> sub_4_cast_fp16 = sub(x = reshape_8_cast_fp16, y = reduce_mean_6_to_fp16)[name = tensor<string, []>("sub_4_cast_fp16")];
            tensor<fp16, [1, 32, 32, ?]> square_2_cast_fp16 = square(x = sub_4_cast_fp16)[name = tensor<string, []>("square_2_cast_fp16")];
            tensor<string, []> square_2_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("square_2_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<int32, [2]> reduce_mean_8_axes_0 = const()[name = tensor<string, []>("reduce_mean_8_axes_0"), val = tensor<int32, [2]>([2, 3])];
            tensor<bool, []> reduce_mean_8_keep_dims_0 = const()[name = tensor<string, []>("reduce_mean_8_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 32, 32, ?]> square_2_cast_fp16_to_fp32 = cast(dtype = square_2_cast_fp16_to_fp32_dtype_0, x = square_2_cast_fp16)[name = tensor<string, []>("cast_224")];
            tensor<fp32, [1, 32, 1, 1]> reduce_mean_8 = reduce_mean(axes = reduce_mean_8_axes_0, keep_dims = reduce_mean_8_keep_dims_0, x = square_2_cast_fp16_to_fp32)[name = tensor<string, []>("reduce_mean_8")];
            tensor<string, []> reduce_mean_8_to_fp16_dtype_0 = const()[name = tensor<string, []>("reduce_mean_8_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> add_4_y_0_to_fp16 = const()[name = tensor<string, []>("add_4_y_0_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 32, 1, 1]> reduce_mean_8_to_fp16 = cast(dtype = reduce_mean_8_to_fp16_dtype_0, x = reduce_mean_8)[name = tensor<string, []>("cast_223")];
            tensor<fp16, [1, 32, 1, 1]> add_4_cast_fp16 = add(x = reduce_mean_8_to_fp16, y = add_4_y_0_to_fp16)[name = tensor<string, []>("add_4_cast_fp16")];
            tensor<fp16, [1, 32, 1, 1]> sqrt_2_cast_fp16 = sqrt(x = add_4_cast_fp16)[name = tensor<string, []>("sqrt_2_cast_fp16")];
            tensor<fp16, [1, 32, 32, ?]> real_div_2_cast_fp16 = real_div(x = sub_4_cast_fp16, y = sqrt_2_cast_fp16)[name = tensor<string, []>("real_div_2_cast_fp16")];
            tensor<fp16, [1, 1024, ?]> reshape_9_cast_fp16 = reshape(shape = shape_2_cast_fp16, x = real_div_2_cast_fp16)[name = tensor<string, []>("reshape_9_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> reshape_10_to_fp16 = const()[name = tensor<string, []>("reshape_10_to_fp16"), val = tensor<fp16, [1, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(32568448)))];
            tensor<fp16, [1, 1024, ?]> mul_2_cast_fp16 = mul(x = reshape_9_cast_fp16, y = reshape_10_to_fp16)[name = tensor<string, []>("mul_2_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> reshape_11_to_fp16 = const()[name = tensor<string, []>("reshape_11_to_fp16"), val = tensor<fp16, [1, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(32570560)))];
            tensor<fp16, [1, 1024, ?]> add_5_cast_fp16 = add(x = mul_2_cast_fp16, y = reshape_11_to_fp16)[name = tensor<string, []>("add_5_cast_fp16")];
            tensor<fp16, [1, 1024, ?]> input_21_cast_fp16 = silu(x = add_5_cast_fp16)[name = tensor<string, []>("input_21_cast_fp16")];
            tensor<string, []> input_23_pad_type_0 = const()[name = tensor<string, []>("input_23_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [2]> input_23_pad_0 = const()[name = tensor<string, []>("input_23_pad_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [1]> input_23_strides_0 = const()[name = tensor<string, []>("input_23_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> input_23_dilations_0 = const()[name = tensor<string, []>("input_23_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> input_23_groups_0 = const()[name = tensor<string, []>("input_23_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 3]> prior_net_1_conv1_weight_to_fp16 = const()[name = tensor<string, []>("prior_net_1_conv1_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 3]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(32572672)))];
            tensor<fp16, [1024]> prior_net_1_conv1_bias_to_fp16 = const()[name = tensor<string, []>("prior_net_1_conv1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(38864192)))];
            tensor<fp16, [1, 1024, ?]> input_23_cast_fp16 = conv(bias = prior_net_1_conv1_bias_to_fp16, dilations = input_23_dilations_0, groups = input_23_groups_0, pad = input_23_pad_0, pad_type = input_23_pad_type_0, strides = input_23_strides_0, weight = prior_net_1_conv1_weight_to_fp16, x = input_21_cast_fp16)[name = tensor<string, []>("input_23_cast_fp16")];
            tensor<int32, [3]> shape_3_cast_fp16 = shape(x = input_23_cast_fp16)[name = tensor<string, []>("shape_3_cast_fp16")];
            tensor<int32, [4]> concat_3x = const()[name = tensor<string, []>("concat_3x"), val = tensor<int32, [4]>([1, 32, 32, -1])];
            tensor<fp16, [1, 32, 32, ?]> reshape_12_cast_fp16 = reshape(shape = concat_3x, x = input_23_cast_fp16)[name = tensor<string, []>("reshape_12_cast_fp16")];
            tensor<string, []> reshape_12_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("reshape_12_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<int32, [2]> reduce_mean_9_axes_0 = const()[name = tensor<string, []>("reduce_mean_9_axes_0"), val = tensor<int32, [2]>([2, 3])];
            tensor<bool, []> reduce_mean_9_keep_dims_0 = const()[name = tensor<string, []>("reduce_mean_9_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 32, 32, ?]> reshape_12_cast_fp16_to_fp32 = cast(dtype = reshape_12_cast_fp16_to_fp32_dtype_0, x = reshape_12_cast_fp16)[name = tensor<string, []>("cast_222")];
            tensor<fp32, [1, 32, 1, 1]> reduce_mean_9 = reduce_mean(axes = reduce_mean_9_axes_0, keep_dims = reduce_mean_9_keep_dims_0, x = reshape_12_cast_fp16_to_fp32)[name = tensor<string, []>("reduce_mean_9")];
            tensor<string, []> reduce_mean_9_to_fp16_dtype_0 = const()[name = tensor<string, []>("reduce_mean_9_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 32, 1, 1]> reduce_mean_9_to_fp16 = cast(dtype = reduce_mean_9_to_fp16_dtype_0, x = reduce_mean_9)[name = tensor<string, []>("cast_221")];
            tensor<fp16, [1, 32, 32, ?]> sub_6_cast_fp16 = sub(x = reshape_12_cast_fp16, y = reduce_mean_9_to_fp16)[name = tensor<string, []>("sub_6_cast_fp16")];
            tensor<fp16, [1, 32, 32, ?]> square_3_cast_fp16 = square(x = sub_6_cast_fp16)[name = tensor<string, []>("square_3_cast_fp16")];
            tensor<string, []> square_3_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("square_3_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<int32, [2]> reduce_mean_11_axes_0 = const()[name = tensor<string, []>("reduce_mean_11_axes_0"), val = tensor<int32, [2]>([2, 3])];
            tensor<bool, []> reduce_mean_11_keep_dims_0 = const()[name = tensor<string, []>("reduce_mean_11_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 32, 32, ?]> square_3_cast_fp16_to_fp32 = cast(dtype = square_3_cast_fp16_to_fp32_dtype_0, x = square_3_cast_fp16)[name = tensor<string, []>("cast_220")];
            tensor<fp32, [1, 32, 1, 1]> reduce_mean_11 = reduce_mean(axes = reduce_mean_11_axes_0, keep_dims = reduce_mean_11_keep_dims_0, x = square_3_cast_fp16_to_fp32)[name = tensor<string, []>("reduce_mean_11")];
            tensor<string, []> reduce_mean_11_to_fp16_dtype_0 = const()[name = tensor<string, []>("reduce_mean_11_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> add_6_y_0_to_fp16 = const()[name = tensor<string, []>("add_6_y_0_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 32, 1, 1]> reduce_mean_11_to_fp16 = cast(dtype = reduce_mean_11_to_fp16_dtype_0, x = reduce_mean_11)[name = tensor<string, []>("cast_219")];
            tensor<fp16, [1, 32, 1, 1]> add_6_cast_fp16 = add(x = reduce_mean_11_to_fp16, y = add_6_y_0_to_fp16)[name = tensor<string, []>("add_6_cast_fp16")];
            tensor<fp16, [1, 32, 1, 1]> sqrt_3_cast_fp16 = sqrt(x = add_6_cast_fp16)[name = tensor<string, []>("sqrt_3_cast_fp16")];
            tensor<fp16, [1, 32, 32, ?]> real_div_3_cast_fp16 = real_div(x = sub_6_cast_fp16, y = sqrt_3_cast_fp16)[name = tensor<string, []>("real_div_3_cast_fp16")];
            tensor<fp16, [1, 1024, ?]> reshape_13_cast_fp16 = reshape(shape = shape_3_cast_fp16, x = real_div_3_cast_fp16)[name = tensor<string, []>("reshape_13_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> reshape_14_to_fp16 = const()[name = tensor<string, []>("reshape_14_to_fp16"), val = tensor<fp16, [1, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(38866304)))];
            tensor<fp16, [1, 1024, ?]> mul_3_cast_fp16 = mul(x = reshape_13_cast_fp16, y = reshape_14_to_fp16)[name = tensor<string, []>("mul_3_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> reshape_15_to_fp16 = const()[name = tensor<string, []>("reshape_15_to_fp16"), val = tensor<fp16, [1, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(38868416)))];
            tensor<fp16, [1, 1024, ?]> add_7_cast_fp16 = add(x = mul_3_cast_fp16, y = reshape_15_to_fp16)[name = tensor<string, []>("add_7_cast_fp16")];
            tensor<fp16, [1, 1024, ?]> input_25_cast_fp16 = silu(x = add_7_cast_fp16)[name = tensor<string, []>("input_25_cast_fp16")];
            tensor<string, []> h_3_pad_type_0 = const()[name = tensor<string, []>("h_3_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [2]> h_3_pad_0 = const()[name = tensor<string, []>("h_3_pad_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [1]> h_3_strides_0 = const()[name = tensor<string, []>("h_3_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> h_3_dilations_0 = const()[name = tensor<string, []>("h_3_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> h_3_groups_0 = const()[name = tensor<string, []>("h_3_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 3]> prior_net_1_conv2_weight_to_fp16 = const()[name = tensor<string, []>("prior_net_1_conv2_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 3]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(38870528)))];
            tensor<fp16, [1024]> prior_net_1_conv2_bias_to_fp16 = const()[name = tensor<string, []>("prior_net_1_conv2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45162048)))];
            tensor<fp16, [1, 1024, ?]> h_3_cast_fp16 = conv(bias = prior_net_1_conv2_bias_to_fp16, dilations = h_3_dilations_0, groups = h_3_groups_0, pad = h_3_pad_0, pad_type = h_3_pad_type_0, strides = h_3_strides_0, weight = prior_net_1_conv2_weight_to_fp16, x = input_25_cast_fp16)[name = tensor<string, []>("h_3_cast_fp16")];
            tensor<fp16, [1, 1024, ?]> x_11_cast_fp16 = add(x = input_19_cast_fp16, y = h_3_cast_fp16)[name = tensor<string, []>("x_11_cast_fp16")];
            tensor<int32, [3]> x_13_perm_0 = const()[name = tensor<string, []>("x_13_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> x_13_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_13_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_145_promoted = const()[name = tensor<string, []>("op_145_promoted"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp16, [1, ?, 1024]> x_13_cast_fp16 = transpose(perm = x_13_perm_0, x = x_11_cast_fp16)[name = tensor<string, []>("transpose_51")];
            tensor<fp32, [1, ?, 1024]> x_13_cast_fp16_to_fp32 = cast(dtype = x_13_cast_fp16_to_fp32_dtype_0, x = x_13_cast_fp16)[name = tensor<string, []>("cast_218")];
            tensor<fp32, [1, ?, 1024]> var_155 = pow(x = x_13_cast_fp16_to_fp32, y = var_145_promoted)[name = tensor<string, []>("op_155")];
            tensor<int32, [1]> norm_x_1_axes_0 = const()[name = tensor<string, []>("norm_x_1_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> norm_x_1_keep_dims_0 = const()[name = tensor<string, []>("norm_x_1_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, ?, 1]> norm_x_1 = reduce_mean(axes = norm_x_1_axes_0, keep_dims = norm_x_1_keep_dims_0, x = var_155)[name = tensor<string, []>("norm_x_1")];
            tensor<string, []> norm_x_1_to_fp16_dtype_0 = const()[name = tensor<string, []>("norm_x_1_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_158_to_fp16 = const()[name = tensor<string, []>("op_158_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> norm_x_1_to_fp16 = cast(dtype = norm_x_1_to_fp16_dtype_0, x = norm_x_1)[name = tensor<string, []>("cast_217")];
            tensor<fp16, [1, ?, 1]> var_159_cast_fp16 = add(x = norm_x_1_to_fp16, y = var_158_to_fp16)[name = tensor<string, []>("op_159_cast_fp16")];
            tensor<string, []> var_159_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_159_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_160_epsilon_0 = const()[name = tensor<string, []>("op_160_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, ?, 1]> var_159_cast_fp16_to_fp32 = cast(dtype = var_159_cast_fp16_to_fp32_dtype_0, x = var_159_cast_fp16)[name = tensor<string, []>("cast_216")];
            tensor<fp32, [1, ?, 1]> var_160 = rsqrt(epsilon = var_160_epsilon_0, x = var_159_cast_fp16_to_fp32)[name = tensor<string, []>("op_160")];
            tensor<string, []> var_160_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_160_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, ?, 1]> var_160_to_fp16 = cast(dtype = var_160_to_fp16_dtype_0, x = var_160)[name = tensor<string, []>("cast_215")];
            tensor<fp16, [1, ?, 1024]> var_161_cast_fp16 = mul(x = x_13_cast_fp16, y = var_160_to_fp16)[name = tensor<string, []>("op_161_cast_fp16")];
            tensor<fp16, [1024]> blocks_0_att_norm_weight_to_fp16 = const()[name = tensor<string, []>("blocks_0_att_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45164160)))];
            tensor<fp16, [1, ?, 1024]> input_29_cast_fp16 = mul(x = var_161_cast_fp16, y = blocks_0_att_norm_weight_to_fp16)[name = tensor<string, []>("input_29_cast_fp16")];
            tensor<fp16, [3072, 1024]> blocks_0_att_c_attn_weight_to_fp16 = const()[name = tensor<string, []>("blocks_0_att_c_attn_weight_to_fp16"), val = tensor<fp16, [3072, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45166272)))];
            tensor<fp16, [3072]> linear_2_bias_0_to_fp16 = const()[name = tensor<string, []>("linear_2_bias_0_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(51457792)))];
            tensor<fp16, [1, ?, 3072]> linear_2_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = blocks_0_att_c_attn_weight_to_fp16, x = input_29_cast_fp16)[name = tensor<string, []>("linear_2_cast_fp16")];
            tensor<int32, [3]> tile_0 = const()[name = tensor<string, []>("tile_0"), val = tensor<int32, [3]>([1024, 1024, 1024])];
            tensor<int32, []> var_169_axis_0 = const()[name = tensor<string, []>("op_169_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, ?, 1024]> var_169_cast_fp16_0, tensor<fp16, [1, ?, 1024]> var_169_cast_fp16_1, tensor<fp16, [1, ?, 1024]> var_169_cast_fp16_2 = split(axis = var_169_axis_0, split_sizes = tile_0, x = linear_2_cast_fp16)[name = tensor<string, []>("op_169_cast_fp16")];
            tensor<int32, [4]> var_173 = const()[name = tensor<string, []>("op_173"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_174_cast_fp16 = reshape(shape = var_173, x = var_169_cast_fp16_0)[name = tensor<string, []>("op_174_cast_fp16")];
            tensor<int32, [4]> x_15_perm_0 = const()[name = tensor<string, []>("x_15_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_176 = const()[name = tensor<string, []>("op_176"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_177_cast_fp16 = reshape(shape = var_176, x = var_169_cast_fp16_1)[name = tensor<string, []>("op_177_cast_fp16")];
            tensor<int32, [4]> x_17_perm_0 = const()[name = tensor<string, []>("x_17_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_179 = const()[name = tensor<string, []>("op_179"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_180_cast_fp16 = reshape(shape = var_179, x = var_169_cast_fp16_2)[name = tensor<string, []>("op_180_cast_fp16")];
            tensor<int32, [4]> v_3_perm_0 = const()[name = tensor<string, []>("v_3_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> x1_1_begin_0 = const()[name = tensor<string, []>("x1_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_1_end_0 = const()[name = tensor<string, []>("x1_1_end_0"), val = tensor<int32, [4]>([1, 16, 0, 64])];
            tensor<bool, [4]> x1_1_end_mask_0 = const()[name = tensor<string, []>("x1_1_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> x1_1_stride_0 = const()[name = tensor<string, []>("x1_1_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 16, ?, 64]> x_15_cast_fp16 = transpose(perm = x_15_perm_0, x = var_174_cast_fp16)[name = tensor<string, []>("transpose_50")];
            tensor<fp16, [1, 16, ?, 32]> x1_1_cast_fp16 = slice_by_index(begin = x1_1_begin_0, end = x1_1_end_0, end_mask = x1_1_end_mask_0, stride = x1_1_stride_0, x = x_15_cast_fp16)[name = tensor<string, []>("x1_1_cast_fp16")];
            tensor<int32, [4]> x2_1_begin_0 = const()[name = tensor<string, []>("x2_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 1])];
            tensor<int32, [4]> x2_1_end_0 = const()[name = tensor<string, []>("x2_1_end_0"), val = tensor<int32, [4]>([1, 16, 0, 64])];
            tensor<bool, [4]> x2_1_end_mask_0 = const()[name = tensor<string, []>("x2_1_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> x2_1_stride_0 = const()[name = tensor<string, []>("x2_1_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 16, ?, 32]> x2_1_cast_fp16 = slice_by_index(begin = x2_1_begin_0, end = x2_1_end_0, end_mask = x2_1_end_mask_0, stride = x2_1_stride_0, x = x_15_cast_fp16)[name = tensor<string, []>("x2_1_cast_fp16")];
            tensor<fp16, [1, 16, 1, 32]> blocks_0_att_head_cos_to_fp16 = const()[name = tensor<string, []>("blocks_0_att_head_cos_to_fp16"), val = tensor<fp16, [1, 16, 1, 32]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(51464000)))];
            tensor<fp16, [1, 16, ?, 32]> var_184_cast_fp16 = mul(x = x1_1_cast_fp16, y = blocks_0_att_head_cos_to_fp16)[name = tensor<string, []>("op_184_cast_fp16")];
            tensor<fp16, [1, 16, 1, 32]> blocks_0_att_head_sin_to_fp16 = const()[name = tensor<string, []>("blocks_0_att_head_sin_to_fp16"), val = tensor<fp16, [1, 16, 1, 32]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(51465088)))];
            tensor<fp16, [1, 16, ?, 32]> var_185_cast_fp16 = mul(x = x2_1_cast_fp16, y = blocks_0_att_head_sin_to_fp16)[name = tensor<string, []>("op_185_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_186_cast_fp16 = sub(x = var_184_cast_fp16, y = var_185_cast_fp16)[name = tensor<string, []>("op_186_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_187_cast_fp16 = mul(x = x2_1_cast_fp16, y = blocks_0_att_head_cos_to_fp16)[name = tensor<string, []>("op_187_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_188_cast_fp16 = mul(x = x1_1_cast_fp16, y = blocks_0_att_head_sin_to_fp16)[name = tensor<string, []>("op_188_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_189_cast_fp16 = add(x = var_187_cast_fp16, y = var_188_cast_fp16)[name = tensor<string, []>("op_189_cast_fp16")];
            tensor<int32, []> out_1_axis_0 = const()[name = tensor<string, []>("out_1_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 16, ?, 32, 2]> out_1_cast_fp16 = stack(axis = out_1_axis_0, values = (var_186_cast_fp16, var_189_cast_fp16))[name = tensor<string, []>("out_1_cast_fp16")];
            tensor<int32, [4]> concat_4x = const()[name = tensor<string, []>("concat_4x"), val = tensor<int32, [4]>([1, 16, -1, 64])];
            tensor<fp16, [1, 16, ?, 64]> q_3_cast_fp16 = reshape(shape = concat_4x, x = out_1_cast_fp16)[name = tensor<string, []>("q_3_cast_fp16")];
            tensor<int32, [4]> x1_3_begin_0 = const()[name = tensor<string, []>("x1_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_3_end_0 = const()[name = tensor<string, []>("x1_3_end_0"), val = tensor<int32, [4]>([1, 16, 0, 64])];
            tensor<bool, [4]> x1_3_end_mask_0 = const()[name = tensor<string, []>("x1_3_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> x1_3_stride_0 = const()[name = tensor<string, []>("x1_3_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 16, ?, 64]> x_17_cast_fp16 = transpose(perm = x_17_perm_0, x = var_177_cast_fp16)[name = tensor<string, []>("transpose_49")];
            tensor<fp16, [1, 16, ?, 32]> x1_3_cast_fp16 = slice_by_index(begin = x1_3_begin_0, end = x1_3_end_0, end_mask = x1_3_end_mask_0, stride = x1_3_stride_0, x = x_17_cast_fp16)[name = tensor<string, []>("x1_3_cast_fp16")];
            tensor<int32, [4]> x2_3_begin_0 = const()[name = tensor<string, []>("x2_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 1])];
            tensor<int32, [4]> x2_3_end_0 = const()[name = tensor<string, []>("x2_3_end_0"), val = tensor<int32, [4]>([1, 16, 0, 64])];
            tensor<bool, [4]> x2_3_end_mask_0 = const()[name = tensor<string, []>("x2_3_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> x2_3_stride_0 = const()[name = tensor<string, []>("x2_3_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 16, ?, 32]> x2_3_cast_fp16 = slice_by_index(begin = x2_3_begin_0, end = x2_3_end_0, end_mask = x2_3_end_mask_0, stride = x2_3_stride_0, x = x_17_cast_fp16)[name = tensor<string, []>("x2_3_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_195_cast_fp16 = mul(x = x1_3_cast_fp16, y = blocks_0_att_head_cos_to_fp16)[name = tensor<string, []>("op_195_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_196_cast_fp16 = mul(x = x2_3_cast_fp16, y = blocks_0_att_head_sin_to_fp16)[name = tensor<string, []>("op_196_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_197_cast_fp16 = sub(x = var_195_cast_fp16, y = var_196_cast_fp16)[name = tensor<string, []>("op_197_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_198_cast_fp16 = mul(x = x2_3_cast_fp16, y = blocks_0_att_head_cos_to_fp16)[name = tensor<string, []>("op_198_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_199_cast_fp16 = mul(x = x1_3_cast_fp16, y = blocks_0_att_head_sin_to_fp16)[name = tensor<string, []>("op_199_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_200_cast_fp16 = add(x = var_198_cast_fp16, y = var_199_cast_fp16)[name = tensor<string, []>("op_200_cast_fp16")];
            tensor<int32, []> out_3_axis_0 = const()[name = tensor<string, []>("out_3_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 16, ?, 32, 2]> out_3_cast_fp16 = stack(axis = out_3_axis_0, values = (var_197_cast_fp16, var_200_cast_fp16))[name = tensor<string, []>("out_3_cast_fp16")];
            tensor<int32, [4]> concat_5x = const()[name = tensor<string, []>("concat_5x"), val = tensor<int32, [4]>([1, 16, -1, 64])];
            tensor<fp16, [1, 16, ?, 64]> k_3_cast_fp16 = reshape(shape = concat_5x, x = out_3_cast_fp16)[name = tensor<string, []>("k_3_cast_fp16")];
            tensor<fp16, []> mul_6_y_0_to_fp16 = const()[name = tensor<string, []>("mul_6_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 16, ?, 64]> mul_6_cast_fp16 = mul(x = q_3_cast_fp16, y = mul_6_y_0_to_fp16)[name = tensor<string, []>("mul_6_cast_fp16")];
            tensor<bool, []> matmul_0_transpose_y_0 = const()[name = tensor<string, []>("matmul_0_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_0_transpose_x_0 = const()[name = tensor<string, []>("matmul_0_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 16, ?, ?]> matmul_0_cast_fp16 = matmul(transpose_x = matmul_0_transpose_x_0, transpose_y = matmul_0_transpose_y_0, x = mul_6_cast_fp16, y = k_3_cast_fp16)[name = tensor<string, []>("matmul_0_cast_fp16")];
            tensor<string, []> matmul_0_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("matmul_0_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<int32, []> softmax_0_axis_0 = const()[name = tensor<string, []>("softmax_0_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp32, [1, 16, ?, ?]> matmul_0_cast_fp16_to_fp32 = cast(dtype = matmul_0_cast_fp16_to_fp32_dtype_0, x = matmul_0_cast_fp16)[name = tensor<string, []>("cast_214")];
            tensor<fp32, [1, 16, ?, ?]> softmax_0 = softmax(axis = softmax_0_axis_0, x = matmul_0_cast_fp16_to_fp32)[name = tensor<string, []>("softmax_0")];
            tensor<bool, []> y_1_transpose_x_0 = const()[name = tensor<string, []>("y_1_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> y_1_transpose_y_0 = const()[name = tensor<string, []>("y_1_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<string, []> softmax_0_to_fp16_dtype_0 = const()[name = tensor<string, []>("softmax_0_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 16, ?, ?]> softmax_0_to_fp16 = cast(dtype = softmax_0_to_fp16_dtype_0, x = softmax_0)[name = tensor<string, []>("cast_213")];
            tensor<fp16, [1, 16, ?, 64]> v_3_cast_fp16 = transpose(perm = v_3_perm_0, x = var_180_cast_fp16)[name = tensor<string, []>("transpose_48")];
            tensor<fp16, [1, 16, ?, 64]> y_1_cast_fp16 = matmul(transpose_x = y_1_transpose_x_0, transpose_y = y_1_transpose_y_0, x = softmax_0_to_fp16, y = v_3_cast_fp16)[name = tensor<string, []>("y_1_cast_fp16")];
            tensor<int32, [4]> var_205_perm_0 = const()[name = tensor<string, []>("op_205_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_206 = const()[name = tensor<string, []>("op_206"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, ?, 16, 64]> var_205_cast_fp16 = transpose(perm = var_205_perm_0, x = y_1_cast_fp16)[name = tensor<string, []>("transpose_47")];
            tensor<fp16, [1, ?, 1024]> input_31_cast_fp16 = reshape(shape = var_206, x = var_205_cast_fp16)[name = tensor<string, []>("input_31_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_0_att_c_proj_weight_to_fp16 = const()[name = tensor<string, []>("blocks_0_att_c_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(51466176)))];
            tensor<fp16, [1024]> linear_3_bias_0_to_fp16 = const()[name = tensor<string, []>("linear_3_bias_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(53563392)))];
            tensor<fp16, [1, ?, 1024]> linear_3_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = blocks_0_att_c_proj_weight_to_fp16, x = input_31_cast_fp16)[name = tensor<string, []>("linear_3_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_19_cast_fp16 = add(x = x_13_cast_fp16, y = linear_3_cast_fp16)[name = tensor<string, []>("x_19_cast_fp16")];
            tensor<string, []> x_19_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_19_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_145_promoted_1 = const()[name = tensor<string, []>("op_145_promoted_1"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, ?, 1024]> x_19_cast_fp16_to_fp32 = cast(dtype = x_19_cast_fp16_to_fp32_dtype_0, x = x_19_cast_fp16)[name = tensor<string, []>("cast_212")];
            tensor<fp32, [1, ?, 1024]> var_212 = pow(x = x_19_cast_fp16_to_fp32, y = var_145_promoted_1)[name = tensor<string, []>("op_212")];
            tensor<int32, [1]> norm_x_3_axes_0 = const()[name = tensor<string, []>("norm_x_3_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> norm_x_3_keep_dims_0 = const()[name = tensor<string, []>("norm_x_3_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, ?, 1]> norm_x_3 = reduce_mean(axes = norm_x_3_axes_0, keep_dims = norm_x_3_keep_dims_0, x = var_212)[name = tensor<string, []>("norm_x_3")];
            tensor<string, []> norm_x_3_to_fp16_dtype_0 = const()[name = tensor<string, []>("norm_x_3_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_215_to_fp16 = const()[name = tensor<string, []>("op_215_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> norm_x_3_to_fp16 = cast(dtype = norm_x_3_to_fp16_dtype_0, x = norm_x_3)[name = tensor<string, []>("cast_211")];
            tensor<fp16, [1, ?, 1]> var_216_cast_fp16 = add(x = norm_x_3_to_fp16, y = var_215_to_fp16)[name = tensor<string, []>("op_216_cast_fp16")];
            tensor<string, []> var_216_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_216_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_217_epsilon_0 = const()[name = tensor<string, []>("op_217_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, ?, 1]> var_216_cast_fp16_to_fp32 = cast(dtype = var_216_cast_fp16_to_fp32_dtype_0, x = var_216_cast_fp16)[name = tensor<string, []>("cast_210")];
            tensor<fp32, [1, ?, 1]> var_217 = rsqrt(epsilon = var_217_epsilon_0, x = var_216_cast_fp16_to_fp32)[name = tensor<string, []>("op_217")];
            tensor<string, []> var_217_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_217_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, ?, 1]> var_217_to_fp16 = cast(dtype = var_217_to_fp16_dtype_0, x = var_217)[name = tensor<string, []>("cast_209")];
            tensor<fp16, [1, ?, 1024]> var_218_cast_fp16 = mul(x = x_19_cast_fp16, y = var_217_to_fp16)[name = tensor<string, []>("op_218_cast_fp16")];
            tensor<fp16, [1024]> blocks_0_ffn_norm_weight_to_fp16 = const()[name = tensor<string, []>("blocks_0_ffn_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(53565504)))];
            tensor<fp16, [1, ?, 1024]> input_33_cast_fp16 = mul(x = var_218_cast_fp16, y = blocks_0_ffn_norm_weight_to_fp16)[name = tensor<string, []>("input_33_cast_fp16")];
            tensor<fp16, [4096, 1024]> blocks_0_mlp_fc1_weight_to_fp16 = const()[name = tensor<string, []>("blocks_0_mlp_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(53567616)))];
            tensor<fp16, [4096]> linear_4_bias_0_to_fp16 = const()[name = tensor<string, []>("linear_4_bias_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61956288)))];
            tensor<fp16, [1, ?, 4096]> linear_4_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_0_mlp_fc1_weight_to_fp16, x = input_33_cast_fp16)[name = tensor<string, []>("linear_4_cast_fp16")];
            tensor<fp16, [1, ?, 4096]> input_37_cast_fp16 = silu(x = linear_4_cast_fp16)[name = tensor<string, []>("input_37_cast_fp16")];
            tensor<fp16, [1024, 4096]> blocks_0_mlp_fc2_weight_to_fp16 = const()[name = tensor<string, []>("blocks_0_mlp_fc2_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61964544)))];
            tensor<fp16, [1, ?, 1024]> linear_5_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = blocks_0_mlp_fc2_weight_to_fp16, x = input_37_cast_fp16)[name = tensor<string, []>("linear_5_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_21_cast_fp16 = add(x = x_19_cast_fp16, y = linear_5_cast_fp16)[name = tensor<string, []>("x_21_cast_fp16")];
            tensor<string, []> x_21_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_21_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_237_promoted = const()[name = tensor<string, []>("op_237_promoted"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, ?, 1024]> x_21_cast_fp16_to_fp32 = cast(dtype = x_21_cast_fp16_to_fp32_dtype_0, x = x_21_cast_fp16)[name = tensor<string, []>("cast_208")];
            tensor<fp32, [1, ?, 1024]> var_247 = pow(x = x_21_cast_fp16_to_fp32, y = var_237_promoted)[name = tensor<string, []>("op_247")];
            tensor<int32, [1]> norm_x_5_axes_0 = const()[name = tensor<string, []>("norm_x_5_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> norm_x_5_keep_dims_0 = const()[name = tensor<string, []>("norm_x_5_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, ?, 1]> norm_x_5 = reduce_mean(axes = norm_x_5_axes_0, keep_dims = norm_x_5_keep_dims_0, x = var_247)[name = tensor<string, []>("norm_x_5")];
            tensor<string, []> norm_x_5_to_fp16_dtype_0 = const()[name = tensor<string, []>("norm_x_5_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_250_to_fp16 = const()[name = tensor<string, []>("op_250_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> norm_x_5_to_fp16 = cast(dtype = norm_x_5_to_fp16_dtype_0, x = norm_x_5)[name = tensor<string, []>("cast_207")];
            tensor<fp16, [1, ?, 1]> var_251_cast_fp16 = add(x = norm_x_5_to_fp16, y = var_250_to_fp16)[name = tensor<string, []>("op_251_cast_fp16")];
            tensor<string, []> var_251_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_251_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_252_epsilon_0 = const()[name = tensor<string, []>("op_252_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, ?, 1]> var_251_cast_fp16_to_fp32 = cast(dtype = var_251_cast_fp16_to_fp32_dtype_0, x = var_251_cast_fp16)[name = tensor<string, []>("cast_206")];
            tensor<fp32, [1, ?, 1]> var_252 = rsqrt(epsilon = var_252_epsilon_0, x = var_251_cast_fp16_to_fp32)[name = tensor<string, []>("op_252")];
            tensor<string, []> var_252_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_252_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, ?, 1]> var_252_to_fp16 = cast(dtype = var_252_to_fp16_dtype_0, x = var_252)[name = tensor<string, []>("cast_205")];
            tensor<fp16, [1, ?, 1024]> var_253_cast_fp16 = mul(x = x_21_cast_fp16, y = var_252_to_fp16)[name = tensor<string, []>("op_253_cast_fp16")];
            tensor<fp16, [1024]> blocks_1_att_norm_weight_to_fp16 = const()[name = tensor<string, []>("blocks_1_att_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(70353216)))];
            tensor<fp16, [1, ?, 1024]> input_39_cast_fp16 = mul(x = var_253_cast_fp16, y = blocks_1_att_norm_weight_to_fp16)[name = tensor<string, []>("input_39_cast_fp16")];
            tensor<fp16, [3072, 1024]> blocks_1_att_c_attn_weight_to_fp16 = const()[name = tensor<string, []>("blocks_1_att_c_attn_weight_to_fp16"), val = tensor<fp16, [3072, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(70355328)))];
            tensor<fp16, [1, ?, 3072]> linear_6_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = blocks_1_att_c_attn_weight_to_fp16, x = input_39_cast_fp16)[name = tensor<string, []>("linear_6_cast_fp16")];
            tensor<int32, [3]> tile_1 = const()[name = tensor<string, []>("tile_1"), val = tensor<int32, [3]>([1024, 1024, 1024])];
            tensor<int32, []> var_261_axis_0 = const()[name = tensor<string, []>("op_261_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, ?, 1024]> var_261_cast_fp16_0, tensor<fp16, [1, ?, 1024]> var_261_cast_fp16_1, tensor<fp16, [1, ?, 1024]> var_261_cast_fp16_2 = split(axis = var_261_axis_0, split_sizes = tile_1, x = linear_6_cast_fp16)[name = tensor<string, []>("op_261_cast_fp16")];
            tensor<int32, [4]> var_265 = const()[name = tensor<string, []>("op_265"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_266_cast_fp16 = reshape(shape = var_265, x = var_261_cast_fp16_0)[name = tensor<string, []>("op_266_cast_fp16")];
            tensor<int32, [4]> x_23_perm_0 = const()[name = tensor<string, []>("x_23_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_268 = const()[name = tensor<string, []>("op_268"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_269_cast_fp16 = reshape(shape = var_268, x = var_261_cast_fp16_1)[name = tensor<string, []>("op_269_cast_fp16")];
            tensor<int32, [4]> x_25_perm_0 = const()[name = tensor<string, []>("x_25_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_271 = const()[name = tensor<string, []>("op_271"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_272_cast_fp16 = reshape(shape = var_271, x = var_261_cast_fp16_2)[name = tensor<string, []>("op_272_cast_fp16")];
            tensor<int32, [4]> v_7_perm_0 = const()[name = tensor<string, []>("v_7_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> x1_5_begin_0 = const()[name = tensor<string, []>("x1_5_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_5_end_0 = const()[name = tensor<string, []>("x1_5_end_0"), val = tensor<int32, [4]>([1, 16, 0, 64])];
            tensor<bool, [4]> x1_5_end_mask_0 = const()[name = tensor<string, []>("x1_5_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> x1_5_stride_0 = const()[name = tensor<string, []>("x1_5_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 16, ?, 64]> x_23_cast_fp16 = transpose(perm = x_23_perm_0, x = var_266_cast_fp16)[name = tensor<string, []>("transpose_46")];
            tensor<fp16, [1, 16, ?, 32]> x1_5_cast_fp16 = slice_by_index(begin = x1_5_begin_0, end = x1_5_end_0, end_mask = x1_5_end_mask_0, stride = x1_5_stride_0, x = x_23_cast_fp16)[name = tensor<string, []>("x1_5_cast_fp16")];
            tensor<int32, [4]> x2_5_begin_0 = const()[name = tensor<string, []>("x2_5_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 1])];
            tensor<int32, [4]> x2_5_end_0 = const()[name = tensor<string, []>("x2_5_end_0"), val = tensor<int32, [4]>([1, 16, 0, 64])];
            tensor<bool, [4]> x2_5_end_mask_0 = const()[name = tensor<string, []>("x2_5_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> x2_5_stride_0 = const()[name = tensor<string, []>("x2_5_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 16, ?, 32]> x2_5_cast_fp16 = slice_by_index(begin = x2_5_begin_0, end = x2_5_end_0, end_mask = x2_5_end_mask_0, stride = x2_5_stride_0, x = x_23_cast_fp16)[name = tensor<string, []>("x2_5_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_276_cast_fp16 = mul(x = x1_5_cast_fp16, y = blocks_0_att_head_cos_to_fp16)[name = tensor<string, []>("op_276_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_277_cast_fp16 = mul(x = x2_5_cast_fp16, y = blocks_0_att_head_sin_to_fp16)[name = tensor<string, []>("op_277_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_278_cast_fp16 = sub(x = var_276_cast_fp16, y = var_277_cast_fp16)[name = tensor<string, []>("op_278_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_279_cast_fp16 = mul(x = x2_5_cast_fp16, y = blocks_0_att_head_cos_to_fp16)[name = tensor<string, []>("op_279_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_280_cast_fp16 = mul(x = x1_5_cast_fp16, y = blocks_0_att_head_sin_to_fp16)[name = tensor<string, []>("op_280_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_281_cast_fp16 = add(x = var_279_cast_fp16, y = var_280_cast_fp16)[name = tensor<string, []>("op_281_cast_fp16")];
            tensor<int32, []> out_5_axis_0 = const()[name = tensor<string, []>("out_5_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 16, ?, 32, 2]> out_5_cast_fp16 = stack(axis = out_5_axis_0, values = (var_278_cast_fp16, var_281_cast_fp16))[name = tensor<string, []>("out_5_cast_fp16")];
            tensor<int32, [4]> concat_6x = const()[name = tensor<string, []>("concat_6x"), val = tensor<int32, [4]>([1, 16, -1, 64])];
            tensor<fp16, [1, 16, ?, 64]> q_7_cast_fp16 = reshape(shape = concat_6x, x = out_5_cast_fp16)[name = tensor<string, []>("q_7_cast_fp16")];
            tensor<int32, [4]> x1_7_begin_0 = const()[name = tensor<string, []>("x1_7_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_7_end_0 = const()[name = tensor<string, []>("x1_7_end_0"), val = tensor<int32, [4]>([1, 16, 0, 64])];
            tensor<bool, [4]> x1_7_end_mask_0 = const()[name = tensor<string, []>("x1_7_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> x1_7_stride_0 = const()[name = tensor<string, []>("x1_7_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 16, ?, 64]> x_25_cast_fp16 = transpose(perm = x_25_perm_0, x = var_269_cast_fp16)[name = tensor<string, []>("transpose_45")];
            tensor<fp16, [1, 16, ?, 32]> x1_7_cast_fp16 = slice_by_index(begin = x1_7_begin_0, end = x1_7_end_0, end_mask = x1_7_end_mask_0, stride = x1_7_stride_0, x = x_25_cast_fp16)[name = tensor<string, []>("x1_7_cast_fp16")];
            tensor<int32, [4]> x2_7_begin_0 = const()[name = tensor<string, []>("x2_7_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 1])];
            tensor<int32, [4]> x2_7_end_0 = const()[name = tensor<string, []>("x2_7_end_0"), val = tensor<int32, [4]>([1, 16, 0, 64])];
            tensor<bool, [4]> x2_7_end_mask_0 = const()[name = tensor<string, []>("x2_7_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> x2_7_stride_0 = const()[name = tensor<string, []>("x2_7_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 16, ?, 32]> x2_7_cast_fp16 = slice_by_index(begin = x2_7_begin_0, end = x2_7_end_0, end_mask = x2_7_end_mask_0, stride = x2_7_stride_0, x = x_25_cast_fp16)[name = tensor<string, []>("x2_7_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_287_cast_fp16 = mul(x = x1_7_cast_fp16, y = blocks_0_att_head_cos_to_fp16)[name = tensor<string, []>("op_287_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_288_cast_fp16 = mul(x = x2_7_cast_fp16, y = blocks_0_att_head_sin_to_fp16)[name = tensor<string, []>("op_288_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_289_cast_fp16 = sub(x = var_287_cast_fp16, y = var_288_cast_fp16)[name = tensor<string, []>("op_289_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_290_cast_fp16 = mul(x = x2_7_cast_fp16, y = blocks_0_att_head_cos_to_fp16)[name = tensor<string, []>("op_290_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_291_cast_fp16 = mul(x = x1_7_cast_fp16, y = blocks_0_att_head_sin_to_fp16)[name = tensor<string, []>("op_291_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_292_cast_fp16 = add(x = var_290_cast_fp16, y = var_291_cast_fp16)[name = tensor<string, []>("op_292_cast_fp16")];
            tensor<int32, []> out_7_axis_0 = const()[name = tensor<string, []>("out_7_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 16, ?, 32, 2]> out_7_cast_fp16 = stack(axis = out_7_axis_0, values = (var_289_cast_fp16, var_292_cast_fp16))[name = tensor<string, []>("out_7_cast_fp16")];
            tensor<int32, [4]> concat_7x = const()[name = tensor<string, []>("concat_7x"), val = tensor<int32, [4]>([1, 16, -1, 64])];
            tensor<fp16, [1, 16, ?, 64]> k_7_cast_fp16 = reshape(shape = concat_7x, x = out_7_cast_fp16)[name = tensor<string, []>("k_7_cast_fp16")];
            tensor<fp16, []> mul_9_y_0_to_fp16 = const()[name = tensor<string, []>("mul_9_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 16, ?, 64]> mul_9_cast_fp16 = mul(x = q_7_cast_fp16, y = mul_9_y_0_to_fp16)[name = tensor<string, []>("mul_9_cast_fp16")];
            tensor<bool, []> matmul_1_transpose_y_0 = const()[name = tensor<string, []>("matmul_1_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_1_transpose_x_0 = const()[name = tensor<string, []>("matmul_1_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 16, ?, ?]> matmul_1_cast_fp16 = matmul(transpose_x = matmul_1_transpose_x_0, transpose_y = matmul_1_transpose_y_0, x = mul_9_cast_fp16, y = k_7_cast_fp16)[name = tensor<string, []>("matmul_1_cast_fp16")];
            tensor<string, []> matmul_1_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("matmul_1_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<int32, []> softmax_1_axis_0 = const()[name = tensor<string, []>("softmax_1_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp32, [1, 16, ?, ?]> matmul_1_cast_fp16_to_fp32 = cast(dtype = matmul_1_cast_fp16_to_fp32_dtype_0, x = matmul_1_cast_fp16)[name = tensor<string, []>("cast_204")];
            tensor<fp32, [1, 16, ?, ?]> softmax_1 = softmax(axis = softmax_1_axis_0, x = matmul_1_cast_fp16_to_fp32)[name = tensor<string, []>("softmax_1")];
            tensor<bool, []> y_3_transpose_x_0 = const()[name = tensor<string, []>("y_3_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> y_3_transpose_y_0 = const()[name = tensor<string, []>("y_3_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<string, []> softmax_1_to_fp16_dtype_0 = const()[name = tensor<string, []>("softmax_1_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 16, ?, ?]> softmax_1_to_fp16 = cast(dtype = softmax_1_to_fp16_dtype_0, x = softmax_1)[name = tensor<string, []>("cast_203")];
            tensor<fp16, [1, 16, ?, 64]> v_7_cast_fp16 = transpose(perm = v_7_perm_0, x = var_272_cast_fp16)[name = tensor<string, []>("transpose_44")];
            tensor<fp16, [1, 16, ?, 64]> y_3_cast_fp16 = matmul(transpose_x = y_3_transpose_x_0, transpose_y = y_3_transpose_y_0, x = softmax_1_to_fp16, y = v_7_cast_fp16)[name = tensor<string, []>("y_3_cast_fp16")];
            tensor<int32, [4]> var_297_perm_0 = const()[name = tensor<string, []>("op_297_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_298 = const()[name = tensor<string, []>("op_298"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, ?, 16, 64]> var_297_cast_fp16 = transpose(perm = var_297_perm_0, x = y_3_cast_fp16)[name = tensor<string, []>("transpose_43")];
            tensor<fp16, [1, ?, 1024]> input_41_cast_fp16 = reshape(shape = var_298, x = var_297_cast_fp16)[name = tensor<string, []>("input_41_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_1_att_c_proj_weight_to_fp16 = const()[name = tensor<string, []>("blocks_1_att_c_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(76646848)))];
            tensor<fp16, [1, ?, 1024]> linear_7_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = blocks_1_att_c_proj_weight_to_fp16, x = input_41_cast_fp16)[name = tensor<string, []>("linear_7_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_27_cast_fp16 = add(x = x_21_cast_fp16, y = linear_7_cast_fp16)[name = tensor<string, []>("x_27_cast_fp16")];
            tensor<string, []> x_27_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_27_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_237_promoted_1 = const()[name = tensor<string, []>("op_237_promoted_1"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, ?, 1024]> x_27_cast_fp16_to_fp32 = cast(dtype = x_27_cast_fp16_to_fp32_dtype_0, x = x_27_cast_fp16)[name = tensor<string, []>("cast_202")];
            tensor<fp32, [1, ?, 1024]> var_304 = pow(x = x_27_cast_fp16_to_fp32, y = var_237_promoted_1)[name = tensor<string, []>("op_304")];
            tensor<int32, [1]> norm_x_7_axes_0 = const()[name = tensor<string, []>("norm_x_7_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> norm_x_7_keep_dims_0 = const()[name = tensor<string, []>("norm_x_7_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, ?, 1]> norm_x_7 = reduce_mean(axes = norm_x_7_axes_0, keep_dims = norm_x_7_keep_dims_0, x = var_304)[name = tensor<string, []>("norm_x_7")];
            tensor<string, []> norm_x_7_to_fp16_dtype_0 = const()[name = tensor<string, []>("norm_x_7_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_307_to_fp16 = const()[name = tensor<string, []>("op_307_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> norm_x_7_to_fp16 = cast(dtype = norm_x_7_to_fp16_dtype_0, x = norm_x_7)[name = tensor<string, []>("cast_201")];
            tensor<fp16, [1, ?, 1]> var_308_cast_fp16 = add(x = norm_x_7_to_fp16, y = var_307_to_fp16)[name = tensor<string, []>("op_308_cast_fp16")];
            tensor<string, []> var_308_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_308_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_309_epsilon_0 = const()[name = tensor<string, []>("op_309_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, ?, 1]> var_308_cast_fp16_to_fp32 = cast(dtype = var_308_cast_fp16_to_fp32_dtype_0, x = var_308_cast_fp16)[name = tensor<string, []>("cast_200")];
            tensor<fp32, [1, ?, 1]> var_309 = rsqrt(epsilon = var_309_epsilon_0, x = var_308_cast_fp16_to_fp32)[name = tensor<string, []>("op_309")];
            tensor<string, []> var_309_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_309_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, ?, 1]> var_309_to_fp16 = cast(dtype = var_309_to_fp16_dtype_0, x = var_309)[name = tensor<string, []>("cast_199")];
            tensor<fp16, [1, ?, 1024]> var_310_cast_fp16 = mul(x = x_27_cast_fp16, y = var_309_to_fp16)[name = tensor<string, []>("op_310_cast_fp16")];
            tensor<fp16, [1024]> blocks_1_ffn_norm_weight_to_fp16 = const()[name = tensor<string, []>("blocks_1_ffn_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(78744064)))];
            tensor<fp16, [1, ?, 1024]> input_43_cast_fp16 = mul(x = var_310_cast_fp16, y = blocks_1_ffn_norm_weight_to_fp16)[name = tensor<string, []>("input_43_cast_fp16")];
            tensor<fp16, [4096, 1024]> blocks_1_mlp_fc1_weight_to_fp16 = const()[name = tensor<string, []>("blocks_1_mlp_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(78746176)))];
            tensor<fp16, [1, ?, 4096]> linear_8_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_1_mlp_fc1_weight_to_fp16, x = input_43_cast_fp16)[name = tensor<string, []>("linear_8_cast_fp16")];
            tensor<fp16, [1, ?, 4096]> input_47_cast_fp16 = silu(x = linear_8_cast_fp16)[name = tensor<string, []>("input_47_cast_fp16")];
            tensor<fp16, [1024, 4096]> blocks_1_mlp_fc2_weight_to_fp16 = const()[name = tensor<string, []>("blocks_1_mlp_fc2_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(87134848)))];
            tensor<fp16, [1, ?, 1024]> linear_9_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = blocks_1_mlp_fc2_weight_to_fp16, x = input_47_cast_fp16)[name = tensor<string, []>("linear_9_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_29_cast_fp16 = add(x = x_27_cast_fp16, y = linear_9_cast_fp16)[name = tensor<string, []>("x_29_cast_fp16")];
            tensor<string, []> x_29_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_29_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_329_promoted = const()[name = tensor<string, []>("op_329_promoted"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, ?, 1024]> x_29_cast_fp16_to_fp32 = cast(dtype = x_29_cast_fp16_to_fp32_dtype_0, x = x_29_cast_fp16)[name = tensor<string, []>("cast_198")];
            tensor<fp32, [1, ?, 1024]> var_339 = pow(x = x_29_cast_fp16_to_fp32, y = var_329_promoted)[name = tensor<string, []>("op_339")];
            tensor<int32, [1]> norm_x_9_axes_0 = const()[name = tensor<string, []>("norm_x_9_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> norm_x_9_keep_dims_0 = const()[name = tensor<string, []>("norm_x_9_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, ?, 1]> norm_x_9 = reduce_mean(axes = norm_x_9_axes_0, keep_dims = norm_x_9_keep_dims_0, x = var_339)[name = tensor<string, []>("norm_x_9")];
            tensor<string, []> norm_x_9_to_fp16_dtype_0 = const()[name = tensor<string, []>("norm_x_9_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_342_to_fp16 = const()[name = tensor<string, []>("op_342_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> norm_x_9_to_fp16 = cast(dtype = norm_x_9_to_fp16_dtype_0, x = norm_x_9)[name = tensor<string, []>("cast_197")];
            tensor<fp16, [1, ?, 1]> var_343_cast_fp16 = add(x = norm_x_9_to_fp16, y = var_342_to_fp16)[name = tensor<string, []>("op_343_cast_fp16")];
            tensor<string, []> var_343_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_343_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_344_epsilon_0 = const()[name = tensor<string, []>("op_344_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, ?, 1]> var_343_cast_fp16_to_fp32 = cast(dtype = var_343_cast_fp16_to_fp32_dtype_0, x = var_343_cast_fp16)[name = tensor<string, []>("cast_196")];
            tensor<fp32, [1, ?, 1]> var_344 = rsqrt(epsilon = var_344_epsilon_0, x = var_343_cast_fp16_to_fp32)[name = tensor<string, []>("op_344")];
            tensor<string, []> var_344_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_344_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, ?, 1]> var_344_to_fp16 = cast(dtype = var_344_to_fp16_dtype_0, x = var_344)[name = tensor<string, []>("cast_195")];
            tensor<fp16, [1, ?, 1024]> var_345_cast_fp16 = mul(x = x_29_cast_fp16, y = var_344_to_fp16)[name = tensor<string, []>("op_345_cast_fp16")];
            tensor<fp16, [1024]> blocks_2_att_norm_weight_to_fp16 = const()[name = tensor<string, []>("blocks_2_att_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(95523520)))];
            tensor<fp16, [1, ?, 1024]> input_49_cast_fp16 = mul(x = var_345_cast_fp16, y = blocks_2_att_norm_weight_to_fp16)[name = tensor<string, []>("input_49_cast_fp16")];
            tensor<fp16, [3072, 1024]> blocks_2_att_c_attn_weight_to_fp16 = const()[name = tensor<string, []>("blocks_2_att_c_attn_weight_to_fp16"), val = tensor<fp16, [3072, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(95525632)))];
            tensor<fp16, [1, ?, 3072]> linear_10_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = blocks_2_att_c_attn_weight_to_fp16, x = input_49_cast_fp16)[name = tensor<string, []>("linear_10_cast_fp16")];
            tensor<int32, [3]> tile_2 = const()[name = tensor<string, []>("tile_2"), val = tensor<int32, [3]>([1024, 1024, 1024])];
            tensor<int32, []> var_353_axis_0 = const()[name = tensor<string, []>("op_353_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, ?, 1024]> var_353_cast_fp16_0, tensor<fp16, [1, ?, 1024]> var_353_cast_fp16_1, tensor<fp16, [1, ?, 1024]> var_353_cast_fp16_2 = split(axis = var_353_axis_0, split_sizes = tile_2, x = linear_10_cast_fp16)[name = tensor<string, []>("op_353_cast_fp16")];
            tensor<int32, [4]> var_357 = const()[name = tensor<string, []>("op_357"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_358_cast_fp16 = reshape(shape = var_357, x = var_353_cast_fp16_0)[name = tensor<string, []>("op_358_cast_fp16")];
            tensor<int32, [4]> x_31_perm_0 = const()[name = tensor<string, []>("x_31_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_360 = const()[name = tensor<string, []>("op_360"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_361_cast_fp16 = reshape(shape = var_360, x = var_353_cast_fp16_1)[name = tensor<string, []>("op_361_cast_fp16")];
            tensor<int32, [4]> x_33_perm_0 = const()[name = tensor<string, []>("x_33_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_363 = const()[name = tensor<string, []>("op_363"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_364_cast_fp16 = reshape(shape = var_363, x = var_353_cast_fp16_2)[name = tensor<string, []>("op_364_cast_fp16")];
            tensor<int32, [4]> v_11_perm_0 = const()[name = tensor<string, []>("v_11_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> x1_9_begin_0 = const()[name = tensor<string, []>("x1_9_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_9_end_0 = const()[name = tensor<string, []>("x1_9_end_0"), val = tensor<int32, [4]>([1, 16, 0, 64])];
            tensor<bool, [4]> x1_9_end_mask_0 = const()[name = tensor<string, []>("x1_9_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> x1_9_stride_0 = const()[name = tensor<string, []>("x1_9_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 16, ?, 64]> x_31_cast_fp16 = transpose(perm = x_31_perm_0, x = var_358_cast_fp16)[name = tensor<string, []>("transpose_42")];
            tensor<fp16, [1, 16, ?, 32]> x1_9_cast_fp16 = slice_by_index(begin = x1_9_begin_0, end = x1_9_end_0, end_mask = x1_9_end_mask_0, stride = x1_9_stride_0, x = x_31_cast_fp16)[name = tensor<string, []>("x1_9_cast_fp16")];
            tensor<int32, [4]> x2_9_begin_0 = const()[name = tensor<string, []>("x2_9_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 1])];
            tensor<int32, [4]> x2_9_end_0 = const()[name = tensor<string, []>("x2_9_end_0"), val = tensor<int32, [4]>([1, 16, 0, 64])];
            tensor<bool, [4]> x2_9_end_mask_0 = const()[name = tensor<string, []>("x2_9_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> x2_9_stride_0 = const()[name = tensor<string, []>("x2_9_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 16, ?, 32]> x2_9_cast_fp16 = slice_by_index(begin = x2_9_begin_0, end = x2_9_end_0, end_mask = x2_9_end_mask_0, stride = x2_9_stride_0, x = x_31_cast_fp16)[name = tensor<string, []>("x2_9_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_368_cast_fp16 = mul(x = x1_9_cast_fp16, y = blocks_0_att_head_cos_to_fp16)[name = tensor<string, []>("op_368_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_369_cast_fp16 = mul(x = x2_9_cast_fp16, y = blocks_0_att_head_sin_to_fp16)[name = tensor<string, []>("op_369_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_370_cast_fp16 = sub(x = var_368_cast_fp16, y = var_369_cast_fp16)[name = tensor<string, []>("op_370_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_371_cast_fp16 = mul(x = x2_9_cast_fp16, y = blocks_0_att_head_cos_to_fp16)[name = tensor<string, []>("op_371_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_372_cast_fp16 = mul(x = x1_9_cast_fp16, y = blocks_0_att_head_sin_to_fp16)[name = tensor<string, []>("op_372_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_373_cast_fp16 = add(x = var_371_cast_fp16, y = var_372_cast_fp16)[name = tensor<string, []>("op_373_cast_fp16")];
            tensor<int32, []> out_9_axis_0 = const()[name = tensor<string, []>("out_9_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 16, ?, 32, 2]> out_9_cast_fp16 = stack(axis = out_9_axis_0, values = (var_370_cast_fp16, var_373_cast_fp16))[name = tensor<string, []>("out_9_cast_fp16")];
            tensor<int32, [4]> concat_8x = const()[name = tensor<string, []>("concat_8x"), val = tensor<int32, [4]>([1, 16, -1, 64])];
            tensor<fp16, [1, 16, ?, 64]> q_11_cast_fp16 = reshape(shape = concat_8x, x = out_9_cast_fp16)[name = tensor<string, []>("q_11_cast_fp16")];
            tensor<int32, [4]> x1_11_begin_0 = const()[name = tensor<string, []>("x1_11_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_11_end_0 = const()[name = tensor<string, []>("x1_11_end_0"), val = tensor<int32, [4]>([1, 16, 0, 64])];
            tensor<bool, [4]> x1_11_end_mask_0 = const()[name = tensor<string, []>("x1_11_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> x1_11_stride_0 = const()[name = tensor<string, []>("x1_11_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 16, ?, 64]> x_33_cast_fp16 = transpose(perm = x_33_perm_0, x = var_361_cast_fp16)[name = tensor<string, []>("transpose_41")];
            tensor<fp16, [1, 16, ?, 32]> x1_11_cast_fp16 = slice_by_index(begin = x1_11_begin_0, end = x1_11_end_0, end_mask = x1_11_end_mask_0, stride = x1_11_stride_0, x = x_33_cast_fp16)[name = tensor<string, []>("x1_11_cast_fp16")];
            tensor<int32, [4]> x2_11_begin_0 = const()[name = tensor<string, []>("x2_11_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 1])];
            tensor<int32, [4]> x2_11_end_0 = const()[name = tensor<string, []>("x2_11_end_0"), val = tensor<int32, [4]>([1, 16, 0, 64])];
            tensor<bool, [4]> x2_11_end_mask_0 = const()[name = tensor<string, []>("x2_11_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> x2_11_stride_0 = const()[name = tensor<string, []>("x2_11_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 16, ?, 32]> x2_11_cast_fp16 = slice_by_index(begin = x2_11_begin_0, end = x2_11_end_0, end_mask = x2_11_end_mask_0, stride = x2_11_stride_0, x = x_33_cast_fp16)[name = tensor<string, []>("x2_11_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_379_cast_fp16 = mul(x = x1_11_cast_fp16, y = blocks_0_att_head_cos_to_fp16)[name = tensor<string, []>("op_379_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_380_cast_fp16 = mul(x = x2_11_cast_fp16, y = blocks_0_att_head_sin_to_fp16)[name = tensor<string, []>("op_380_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_381_cast_fp16 = sub(x = var_379_cast_fp16, y = var_380_cast_fp16)[name = tensor<string, []>("op_381_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_382_cast_fp16 = mul(x = x2_11_cast_fp16, y = blocks_0_att_head_cos_to_fp16)[name = tensor<string, []>("op_382_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_383_cast_fp16 = mul(x = x1_11_cast_fp16, y = blocks_0_att_head_sin_to_fp16)[name = tensor<string, []>("op_383_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_384_cast_fp16 = add(x = var_382_cast_fp16, y = var_383_cast_fp16)[name = tensor<string, []>("op_384_cast_fp16")];
            tensor<int32, []> out_11_axis_0 = const()[name = tensor<string, []>("out_11_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 16, ?, 32, 2]> out_11_cast_fp16 = stack(axis = out_11_axis_0, values = (var_381_cast_fp16, var_384_cast_fp16))[name = tensor<string, []>("out_11_cast_fp16")];
            tensor<int32, [4]> concat_9x = const()[name = tensor<string, []>("concat_9x"), val = tensor<int32, [4]>([1, 16, -1, 64])];
            tensor<fp16, [1, 16, ?, 64]> k_11_cast_fp16 = reshape(shape = concat_9x, x = out_11_cast_fp16)[name = tensor<string, []>("k_11_cast_fp16")];
            tensor<fp16, []> mul_12_y_0_to_fp16 = const()[name = tensor<string, []>("mul_12_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 16, ?, 64]> mul_12_cast_fp16 = mul(x = q_11_cast_fp16, y = mul_12_y_0_to_fp16)[name = tensor<string, []>("mul_12_cast_fp16")];
            tensor<bool, []> matmul_2_transpose_y_0 = const()[name = tensor<string, []>("matmul_2_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_2_transpose_x_0 = const()[name = tensor<string, []>("matmul_2_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 16, ?, ?]> matmul_2_cast_fp16 = matmul(transpose_x = matmul_2_transpose_x_0, transpose_y = matmul_2_transpose_y_0, x = mul_12_cast_fp16, y = k_11_cast_fp16)[name = tensor<string, []>("matmul_2_cast_fp16")];
            tensor<string, []> matmul_2_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("matmul_2_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<int32, []> softmax_2_axis_0 = const()[name = tensor<string, []>("softmax_2_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp32, [1, 16, ?, ?]> matmul_2_cast_fp16_to_fp32 = cast(dtype = matmul_2_cast_fp16_to_fp32_dtype_0, x = matmul_2_cast_fp16)[name = tensor<string, []>("cast_194")];
            tensor<fp32, [1, 16, ?, ?]> softmax_2 = softmax(axis = softmax_2_axis_0, x = matmul_2_cast_fp16_to_fp32)[name = tensor<string, []>("softmax_2")];
            tensor<bool, []> y_5_transpose_x_0 = const()[name = tensor<string, []>("y_5_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> y_5_transpose_y_0 = const()[name = tensor<string, []>("y_5_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<string, []> softmax_2_to_fp16_dtype_0 = const()[name = tensor<string, []>("softmax_2_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 16, ?, ?]> softmax_2_to_fp16 = cast(dtype = softmax_2_to_fp16_dtype_0, x = softmax_2)[name = tensor<string, []>("cast_193")];
            tensor<fp16, [1, 16, ?, 64]> v_11_cast_fp16 = transpose(perm = v_11_perm_0, x = var_364_cast_fp16)[name = tensor<string, []>("transpose_40")];
            tensor<fp16, [1, 16, ?, 64]> y_5_cast_fp16 = matmul(transpose_x = y_5_transpose_x_0, transpose_y = y_5_transpose_y_0, x = softmax_2_to_fp16, y = v_11_cast_fp16)[name = tensor<string, []>("y_5_cast_fp16")];
            tensor<int32, [4]> var_389_perm_0 = const()[name = tensor<string, []>("op_389_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_390 = const()[name = tensor<string, []>("op_390"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, ?, 16, 64]> var_389_cast_fp16 = transpose(perm = var_389_perm_0, x = y_5_cast_fp16)[name = tensor<string, []>("transpose_39")];
            tensor<fp16, [1, ?, 1024]> input_51_cast_fp16 = reshape(shape = var_390, x = var_389_cast_fp16)[name = tensor<string, []>("input_51_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_2_att_c_proj_weight_to_fp16 = const()[name = tensor<string, []>("blocks_2_att_c_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(101817152)))];
            tensor<fp16, [1, ?, 1024]> linear_11_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = blocks_2_att_c_proj_weight_to_fp16, x = input_51_cast_fp16)[name = tensor<string, []>("linear_11_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_35_cast_fp16 = add(x = x_29_cast_fp16, y = linear_11_cast_fp16)[name = tensor<string, []>("x_35_cast_fp16")];
            tensor<string, []> x_35_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_35_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_329_promoted_1 = const()[name = tensor<string, []>("op_329_promoted_1"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, ?, 1024]> x_35_cast_fp16_to_fp32 = cast(dtype = x_35_cast_fp16_to_fp32_dtype_0, x = x_35_cast_fp16)[name = tensor<string, []>("cast_192")];
            tensor<fp32, [1, ?, 1024]> var_396 = pow(x = x_35_cast_fp16_to_fp32, y = var_329_promoted_1)[name = tensor<string, []>("op_396")];
            tensor<int32, [1]> norm_x_11_axes_0 = const()[name = tensor<string, []>("norm_x_11_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> norm_x_11_keep_dims_0 = const()[name = tensor<string, []>("norm_x_11_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, ?, 1]> norm_x_11 = reduce_mean(axes = norm_x_11_axes_0, keep_dims = norm_x_11_keep_dims_0, x = var_396)[name = tensor<string, []>("norm_x_11")];
            tensor<string, []> norm_x_11_to_fp16_dtype_0 = const()[name = tensor<string, []>("norm_x_11_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_399_to_fp16 = const()[name = tensor<string, []>("op_399_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> norm_x_11_to_fp16 = cast(dtype = norm_x_11_to_fp16_dtype_0, x = norm_x_11)[name = tensor<string, []>("cast_191")];
            tensor<fp16, [1, ?, 1]> var_400_cast_fp16 = add(x = norm_x_11_to_fp16, y = var_399_to_fp16)[name = tensor<string, []>("op_400_cast_fp16")];
            tensor<string, []> var_400_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_400_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_401_epsilon_0 = const()[name = tensor<string, []>("op_401_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, ?, 1]> var_400_cast_fp16_to_fp32 = cast(dtype = var_400_cast_fp16_to_fp32_dtype_0, x = var_400_cast_fp16)[name = tensor<string, []>("cast_190")];
            tensor<fp32, [1, ?, 1]> var_401 = rsqrt(epsilon = var_401_epsilon_0, x = var_400_cast_fp16_to_fp32)[name = tensor<string, []>("op_401")];
            tensor<string, []> var_401_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_401_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, ?, 1]> var_401_to_fp16 = cast(dtype = var_401_to_fp16_dtype_0, x = var_401)[name = tensor<string, []>("cast_189")];
            tensor<fp16, [1, ?, 1024]> var_402_cast_fp16 = mul(x = x_35_cast_fp16, y = var_401_to_fp16)[name = tensor<string, []>("op_402_cast_fp16")];
            tensor<fp16, [1024]> blocks_2_ffn_norm_weight_to_fp16 = const()[name = tensor<string, []>("blocks_2_ffn_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(103914368)))];
            tensor<fp16, [1, ?, 1024]> input_53_cast_fp16 = mul(x = var_402_cast_fp16, y = blocks_2_ffn_norm_weight_to_fp16)[name = tensor<string, []>("input_53_cast_fp16")];
            tensor<fp16, [4096, 1024]> blocks_2_mlp_fc1_weight_to_fp16 = const()[name = tensor<string, []>("blocks_2_mlp_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(103916480)))];
            tensor<fp16, [1, ?, 4096]> linear_12_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_2_mlp_fc1_weight_to_fp16, x = input_53_cast_fp16)[name = tensor<string, []>("linear_12_cast_fp16")];
            tensor<fp16, [1, ?, 4096]> input_57_cast_fp16 = silu(x = linear_12_cast_fp16)[name = tensor<string, []>("input_57_cast_fp16")];
            tensor<fp16, [1024, 4096]> blocks_2_mlp_fc2_weight_to_fp16 = const()[name = tensor<string, []>("blocks_2_mlp_fc2_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(112305152)))];
            tensor<fp16, [1, ?, 1024]> linear_13_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = blocks_2_mlp_fc2_weight_to_fp16, x = input_57_cast_fp16)[name = tensor<string, []>("linear_13_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_37_cast_fp16 = add(x = x_35_cast_fp16, y = linear_13_cast_fp16)[name = tensor<string, []>("x_37_cast_fp16")];
            tensor<string, []> x_37_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_37_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_421_promoted = const()[name = tensor<string, []>("op_421_promoted"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, ?, 1024]> x_37_cast_fp16_to_fp32 = cast(dtype = x_37_cast_fp16_to_fp32_dtype_0, x = x_37_cast_fp16)[name = tensor<string, []>("cast_188")];
            tensor<fp32, [1, ?, 1024]> var_431 = pow(x = x_37_cast_fp16_to_fp32, y = var_421_promoted)[name = tensor<string, []>("op_431")];
            tensor<int32, [1]> norm_x_13_axes_0 = const()[name = tensor<string, []>("norm_x_13_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> norm_x_13_keep_dims_0 = const()[name = tensor<string, []>("norm_x_13_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, ?, 1]> norm_x_13 = reduce_mean(axes = norm_x_13_axes_0, keep_dims = norm_x_13_keep_dims_0, x = var_431)[name = tensor<string, []>("norm_x_13")];
            tensor<string, []> norm_x_13_to_fp16_dtype_0 = const()[name = tensor<string, []>("norm_x_13_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_434_to_fp16 = const()[name = tensor<string, []>("op_434_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> norm_x_13_to_fp16 = cast(dtype = norm_x_13_to_fp16_dtype_0, x = norm_x_13)[name = tensor<string, []>("cast_187")];
            tensor<fp16, [1, ?, 1]> var_435_cast_fp16 = add(x = norm_x_13_to_fp16, y = var_434_to_fp16)[name = tensor<string, []>("op_435_cast_fp16")];
            tensor<string, []> var_435_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_435_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_436_epsilon_0 = const()[name = tensor<string, []>("op_436_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, ?, 1]> var_435_cast_fp16_to_fp32 = cast(dtype = var_435_cast_fp16_to_fp32_dtype_0, x = var_435_cast_fp16)[name = tensor<string, []>("cast_186")];
            tensor<fp32, [1, ?, 1]> var_436 = rsqrt(epsilon = var_436_epsilon_0, x = var_435_cast_fp16_to_fp32)[name = tensor<string, []>("op_436")];
            tensor<string, []> var_436_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_436_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, ?, 1]> var_436_to_fp16 = cast(dtype = var_436_to_fp16_dtype_0, x = var_436)[name = tensor<string, []>("cast_185")];
            tensor<fp16, [1, ?, 1024]> var_437_cast_fp16 = mul(x = x_37_cast_fp16, y = var_436_to_fp16)[name = tensor<string, []>("op_437_cast_fp16")];
            tensor<fp16, [1024]> blocks_3_att_norm_weight_to_fp16 = const()[name = tensor<string, []>("blocks_3_att_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(120693824)))];
            tensor<fp16, [1, ?, 1024]> input_59_cast_fp16 = mul(x = var_437_cast_fp16, y = blocks_3_att_norm_weight_to_fp16)[name = tensor<string, []>("input_59_cast_fp16")];
            tensor<fp16, [3072, 1024]> blocks_3_att_c_attn_weight_to_fp16 = const()[name = tensor<string, []>("blocks_3_att_c_attn_weight_to_fp16"), val = tensor<fp16, [3072, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(120695936)))];
            tensor<fp16, [1, ?, 3072]> linear_14_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = blocks_3_att_c_attn_weight_to_fp16, x = input_59_cast_fp16)[name = tensor<string, []>("linear_14_cast_fp16")];
            tensor<int32, [3]> tile_3 = const()[name = tensor<string, []>("tile_3"), val = tensor<int32, [3]>([1024, 1024, 1024])];
            tensor<int32, []> var_445_axis_0 = const()[name = tensor<string, []>("op_445_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, ?, 1024]> var_445_cast_fp16_0, tensor<fp16, [1, ?, 1024]> var_445_cast_fp16_1, tensor<fp16, [1, ?, 1024]> var_445_cast_fp16_2 = split(axis = var_445_axis_0, split_sizes = tile_3, x = linear_14_cast_fp16)[name = tensor<string, []>("op_445_cast_fp16")];
            tensor<int32, [4]> var_449 = const()[name = tensor<string, []>("op_449"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_450_cast_fp16 = reshape(shape = var_449, x = var_445_cast_fp16_0)[name = tensor<string, []>("op_450_cast_fp16")];
            tensor<int32, [4]> x_39_perm_0 = const()[name = tensor<string, []>("x_39_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_452 = const()[name = tensor<string, []>("op_452"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_453_cast_fp16 = reshape(shape = var_452, x = var_445_cast_fp16_1)[name = tensor<string, []>("op_453_cast_fp16")];
            tensor<int32, [4]> x_41_perm_0 = const()[name = tensor<string, []>("x_41_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_455 = const()[name = tensor<string, []>("op_455"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_456_cast_fp16 = reshape(shape = var_455, x = var_445_cast_fp16_2)[name = tensor<string, []>("op_456_cast_fp16")];
            tensor<int32, [4]> v_15_perm_0 = const()[name = tensor<string, []>("v_15_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> x1_13_begin_0 = const()[name = tensor<string, []>("x1_13_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_13_end_0 = const()[name = tensor<string, []>("x1_13_end_0"), val = tensor<int32, [4]>([1, 16, 0, 64])];
            tensor<bool, [4]> x1_13_end_mask_0 = const()[name = tensor<string, []>("x1_13_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> x1_13_stride_0 = const()[name = tensor<string, []>("x1_13_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 16, ?, 64]> x_39_cast_fp16 = transpose(perm = x_39_perm_0, x = var_450_cast_fp16)[name = tensor<string, []>("transpose_38")];
            tensor<fp16, [1, 16, ?, 32]> x1_13_cast_fp16 = slice_by_index(begin = x1_13_begin_0, end = x1_13_end_0, end_mask = x1_13_end_mask_0, stride = x1_13_stride_0, x = x_39_cast_fp16)[name = tensor<string, []>("x1_13_cast_fp16")];
            tensor<int32, [4]> x2_13_begin_0 = const()[name = tensor<string, []>("x2_13_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 1])];
            tensor<int32, [4]> x2_13_end_0 = const()[name = tensor<string, []>("x2_13_end_0"), val = tensor<int32, [4]>([1, 16, 0, 64])];
            tensor<bool, [4]> x2_13_end_mask_0 = const()[name = tensor<string, []>("x2_13_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> x2_13_stride_0 = const()[name = tensor<string, []>("x2_13_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 16, ?, 32]> x2_13_cast_fp16 = slice_by_index(begin = x2_13_begin_0, end = x2_13_end_0, end_mask = x2_13_end_mask_0, stride = x2_13_stride_0, x = x_39_cast_fp16)[name = tensor<string, []>("x2_13_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_460_cast_fp16 = mul(x = x1_13_cast_fp16, y = blocks_0_att_head_cos_to_fp16)[name = tensor<string, []>("op_460_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_461_cast_fp16 = mul(x = x2_13_cast_fp16, y = blocks_0_att_head_sin_to_fp16)[name = tensor<string, []>("op_461_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_462_cast_fp16 = sub(x = var_460_cast_fp16, y = var_461_cast_fp16)[name = tensor<string, []>("op_462_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_463_cast_fp16 = mul(x = x2_13_cast_fp16, y = blocks_0_att_head_cos_to_fp16)[name = tensor<string, []>("op_463_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_464_cast_fp16 = mul(x = x1_13_cast_fp16, y = blocks_0_att_head_sin_to_fp16)[name = tensor<string, []>("op_464_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_465_cast_fp16 = add(x = var_463_cast_fp16, y = var_464_cast_fp16)[name = tensor<string, []>("op_465_cast_fp16")];
            tensor<int32, []> out_13_axis_0 = const()[name = tensor<string, []>("out_13_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 16, ?, 32, 2]> out_13_cast_fp16 = stack(axis = out_13_axis_0, values = (var_462_cast_fp16, var_465_cast_fp16))[name = tensor<string, []>("out_13_cast_fp16")];
            tensor<int32, [4]> concat_10x = const()[name = tensor<string, []>("concat_10x"), val = tensor<int32, [4]>([1, 16, -1, 64])];
            tensor<fp16, [1, 16, ?, 64]> q_15_cast_fp16 = reshape(shape = concat_10x, x = out_13_cast_fp16)[name = tensor<string, []>("q_15_cast_fp16")];
            tensor<int32, [4]> x1_15_begin_0 = const()[name = tensor<string, []>("x1_15_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_15_end_0 = const()[name = tensor<string, []>("x1_15_end_0"), val = tensor<int32, [4]>([1, 16, 0, 64])];
            tensor<bool, [4]> x1_15_end_mask_0 = const()[name = tensor<string, []>("x1_15_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> x1_15_stride_0 = const()[name = tensor<string, []>("x1_15_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 16, ?, 64]> x_41_cast_fp16 = transpose(perm = x_41_perm_0, x = var_453_cast_fp16)[name = tensor<string, []>("transpose_37")];
            tensor<fp16, [1, 16, ?, 32]> x1_15_cast_fp16 = slice_by_index(begin = x1_15_begin_0, end = x1_15_end_0, end_mask = x1_15_end_mask_0, stride = x1_15_stride_0, x = x_41_cast_fp16)[name = tensor<string, []>("x1_15_cast_fp16")];
            tensor<int32, [4]> x2_15_begin_0 = const()[name = tensor<string, []>("x2_15_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 1])];
            tensor<int32, [4]> x2_15_end_0 = const()[name = tensor<string, []>("x2_15_end_0"), val = tensor<int32, [4]>([1, 16, 0, 64])];
            tensor<bool, [4]> x2_15_end_mask_0 = const()[name = tensor<string, []>("x2_15_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> x2_15_stride_0 = const()[name = tensor<string, []>("x2_15_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 16, ?, 32]> x2_15_cast_fp16 = slice_by_index(begin = x2_15_begin_0, end = x2_15_end_0, end_mask = x2_15_end_mask_0, stride = x2_15_stride_0, x = x_41_cast_fp16)[name = tensor<string, []>("x2_15_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_471_cast_fp16 = mul(x = x1_15_cast_fp16, y = blocks_0_att_head_cos_to_fp16)[name = tensor<string, []>("op_471_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_472_cast_fp16 = mul(x = x2_15_cast_fp16, y = blocks_0_att_head_sin_to_fp16)[name = tensor<string, []>("op_472_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_473_cast_fp16 = sub(x = var_471_cast_fp16, y = var_472_cast_fp16)[name = tensor<string, []>("op_473_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_474_cast_fp16 = mul(x = x2_15_cast_fp16, y = blocks_0_att_head_cos_to_fp16)[name = tensor<string, []>("op_474_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_475_cast_fp16 = mul(x = x1_15_cast_fp16, y = blocks_0_att_head_sin_to_fp16)[name = tensor<string, []>("op_475_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_476_cast_fp16 = add(x = var_474_cast_fp16, y = var_475_cast_fp16)[name = tensor<string, []>("op_476_cast_fp16")];
            tensor<int32, []> out_15_axis_0 = const()[name = tensor<string, []>("out_15_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 16, ?, 32, 2]> out_15_cast_fp16 = stack(axis = out_15_axis_0, values = (var_473_cast_fp16, var_476_cast_fp16))[name = tensor<string, []>("out_15_cast_fp16")];
            tensor<int32, [4]> concat_11x = const()[name = tensor<string, []>("concat_11x"), val = tensor<int32, [4]>([1, 16, -1, 64])];
            tensor<fp16, [1, 16, ?, 64]> k_15_cast_fp16 = reshape(shape = concat_11x, x = out_15_cast_fp16)[name = tensor<string, []>("k_15_cast_fp16")];
            tensor<fp16, []> mul_15_y_0_to_fp16 = const()[name = tensor<string, []>("mul_15_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 16, ?, 64]> mul_15_cast_fp16 = mul(x = q_15_cast_fp16, y = mul_15_y_0_to_fp16)[name = tensor<string, []>("mul_15_cast_fp16")];
            tensor<bool, []> matmul_3_transpose_y_0 = const()[name = tensor<string, []>("matmul_3_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_3_transpose_x_0 = const()[name = tensor<string, []>("matmul_3_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 16, ?, ?]> matmul_3_cast_fp16 = matmul(transpose_x = matmul_3_transpose_x_0, transpose_y = matmul_3_transpose_y_0, x = mul_15_cast_fp16, y = k_15_cast_fp16)[name = tensor<string, []>("matmul_3_cast_fp16")];
            tensor<string, []> matmul_3_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("matmul_3_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<int32, []> softmax_3_axis_0 = const()[name = tensor<string, []>("softmax_3_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp32, [1, 16, ?, ?]> matmul_3_cast_fp16_to_fp32 = cast(dtype = matmul_3_cast_fp16_to_fp32_dtype_0, x = matmul_3_cast_fp16)[name = tensor<string, []>("cast_184")];
            tensor<fp32, [1, 16, ?, ?]> softmax_3 = softmax(axis = softmax_3_axis_0, x = matmul_3_cast_fp16_to_fp32)[name = tensor<string, []>("softmax_3")];
            tensor<bool, []> y_7_transpose_x_0 = const()[name = tensor<string, []>("y_7_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> y_7_transpose_y_0 = const()[name = tensor<string, []>("y_7_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<string, []> softmax_3_to_fp16_dtype_0 = const()[name = tensor<string, []>("softmax_3_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 16, ?, ?]> softmax_3_to_fp16 = cast(dtype = softmax_3_to_fp16_dtype_0, x = softmax_3)[name = tensor<string, []>("cast_183")];
            tensor<fp16, [1, 16, ?, 64]> v_15_cast_fp16 = transpose(perm = v_15_perm_0, x = var_456_cast_fp16)[name = tensor<string, []>("transpose_36")];
            tensor<fp16, [1, 16, ?, 64]> y_7_cast_fp16 = matmul(transpose_x = y_7_transpose_x_0, transpose_y = y_7_transpose_y_0, x = softmax_3_to_fp16, y = v_15_cast_fp16)[name = tensor<string, []>("y_7_cast_fp16")];
            tensor<int32, [4]> var_481_perm_0 = const()[name = tensor<string, []>("op_481_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_482 = const()[name = tensor<string, []>("op_482"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, ?, 16, 64]> var_481_cast_fp16 = transpose(perm = var_481_perm_0, x = y_7_cast_fp16)[name = tensor<string, []>("transpose_35")];
            tensor<fp16, [1, ?, 1024]> input_61_cast_fp16 = reshape(shape = var_482, x = var_481_cast_fp16)[name = tensor<string, []>("input_61_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_3_att_c_proj_weight_to_fp16 = const()[name = tensor<string, []>("blocks_3_att_c_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(126987456)))];
            tensor<fp16, [1, ?, 1024]> linear_15_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = blocks_3_att_c_proj_weight_to_fp16, x = input_61_cast_fp16)[name = tensor<string, []>("linear_15_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_43_cast_fp16 = add(x = x_37_cast_fp16, y = linear_15_cast_fp16)[name = tensor<string, []>("x_43_cast_fp16")];
            tensor<string, []> x_43_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_43_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_421_promoted_1 = const()[name = tensor<string, []>("op_421_promoted_1"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, ?, 1024]> x_43_cast_fp16_to_fp32 = cast(dtype = x_43_cast_fp16_to_fp32_dtype_0, x = x_43_cast_fp16)[name = tensor<string, []>("cast_182")];
            tensor<fp32, [1, ?, 1024]> var_488 = pow(x = x_43_cast_fp16_to_fp32, y = var_421_promoted_1)[name = tensor<string, []>("op_488")];
            tensor<int32, [1]> norm_x_15_axes_0 = const()[name = tensor<string, []>("norm_x_15_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> norm_x_15_keep_dims_0 = const()[name = tensor<string, []>("norm_x_15_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, ?, 1]> norm_x_15 = reduce_mean(axes = norm_x_15_axes_0, keep_dims = norm_x_15_keep_dims_0, x = var_488)[name = tensor<string, []>("norm_x_15")];
            tensor<string, []> norm_x_15_to_fp16_dtype_0 = const()[name = tensor<string, []>("norm_x_15_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_491_to_fp16 = const()[name = tensor<string, []>("op_491_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> norm_x_15_to_fp16 = cast(dtype = norm_x_15_to_fp16_dtype_0, x = norm_x_15)[name = tensor<string, []>("cast_181")];
            tensor<fp16, [1, ?, 1]> var_492_cast_fp16 = add(x = norm_x_15_to_fp16, y = var_491_to_fp16)[name = tensor<string, []>("op_492_cast_fp16")];
            tensor<string, []> var_492_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_492_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_493_epsilon_0 = const()[name = tensor<string, []>("op_493_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, ?, 1]> var_492_cast_fp16_to_fp32 = cast(dtype = var_492_cast_fp16_to_fp32_dtype_0, x = var_492_cast_fp16)[name = tensor<string, []>("cast_180")];
            tensor<fp32, [1, ?, 1]> var_493 = rsqrt(epsilon = var_493_epsilon_0, x = var_492_cast_fp16_to_fp32)[name = tensor<string, []>("op_493")];
            tensor<string, []> var_493_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_493_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, ?, 1]> var_493_to_fp16 = cast(dtype = var_493_to_fp16_dtype_0, x = var_493)[name = tensor<string, []>("cast_179")];
            tensor<fp16, [1, ?, 1024]> var_494_cast_fp16 = mul(x = x_43_cast_fp16, y = var_493_to_fp16)[name = tensor<string, []>("op_494_cast_fp16")];
            tensor<fp16, [1024]> blocks_3_ffn_norm_weight_to_fp16 = const()[name = tensor<string, []>("blocks_3_ffn_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(129084672)))];
            tensor<fp16, [1, ?, 1024]> input_63_cast_fp16 = mul(x = var_494_cast_fp16, y = blocks_3_ffn_norm_weight_to_fp16)[name = tensor<string, []>("input_63_cast_fp16")];
            tensor<fp16, [4096, 1024]> blocks_3_mlp_fc1_weight_to_fp16 = const()[name = tensor<string, []>("blocks_3_mlp_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(129086784)))];
            tensor<fp16, [1, ?, 4096]> linear_16_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_3_mlp_fc1_weight_to_fp16, x = input_63_cast_fp16)[name = tensor<string, []>("linear_16_cast_fp16")];
            tensor<fp16, [1, ?, 4096]> input_67_cast_fp16 = silu(x = linear_16_cast_fp16)[name = tensor<string, []>("input_67_cast_fp16")];
            tensor<fp16, [1024, 4096]> blocks_3_mlp_fc2_weight_to_fp16 = const()[name = tensor<string, []>("blocks_3_mlp_fc2_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(137475456)))];
            tensor<fp16, [1, ?, 1024]> linear_17_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = blocks_3_mlp_fc2_weight_to_fp16, x = input_67_cast_fp16)[name = tensor<string, []>("linear_17_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_45_cast_fp16 = add(x = x_43_cast_fp16, y = linear_17_cast_fp16)[name = tensor<string, []>("x_45_cast_fp16")];
            tensor<string, []> x_45_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_45_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_513_promoted = const()[name = tensor<string, []>("op_513_promoted"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, ?, 1024]> x_45_cast_fp16_to_fp32 = cast(dtype = x_45_cast_fp16_to_fp32_dtype_0, x = x_45_cast_fp16)[name = tensor<string, []>("cast_178")];
            tensor<fp32, [1, ?, 1024]> var_523 = pow(x = x_45_cast_fp16_to_fp32, y = var_513_promoted)[name = tensor<string, []>("op_523")];
            tensor<int32, [1]> norm_x_17_axes_0 = const()[name = tensor<string, []>("norm_x_17_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> norm_x_17_keep_dims_0 = const()[name = tensor<string, []>("norm_x_17_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, ?, 1]> norm_x_17 = reduce_mean(axes = norm_x_17_axes_0, keep_dims = norm_x_17_keep_dims_0, x = var_523)[name = tensor<string, []>("norm_x_17")];
            tensor<string, []> norm_x_17_to_fp16_dtype_0 = const()[name = tensor<string, []>("norm_x_17_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_526_to_fp16 = const()[name = tensor<string, []>("op_526_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> norm_x_17_to_fp16 = cast(dtype = norm_x_17_to_fp16_dtype_0, x = norm_x_17)[name = tensor<string, []>("cast_177")];
            tensor<fp16, [1, ?, 1]> var_527_cast_fp16 = add(x = norm_x_17_to_fp16, y = var_526_to_fp16)[name = tensor<string, []>("op_527_cast_fp16")];
            tensor<string, []> var_527_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_527_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_528_epsilon_0 = const()[name = tensor<string, []>("op_528_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, ?, 1]> var_527_cast_fp16_to_fp32 = cast(dtype = var_527_cast_fp16_to_fp32_dtype_0, x = var_527_cast_fp16)[name = tensor<string, []>("cast_176")];
            tensor<fp32, [1, ?, 1]> var_528 = rsqrt(epsilon = var_528_epsilon_0, x = var_527_cast_fp16_to_fp32)[name = tensor<string, []>("op_528")];
            tensor<string, []> var_528_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_528_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, ?, 1]> var_528_to_fp16 = cast(dtype = var_528_to_fp16_dtype_0, x = var_528)[name = tensor<string, []>("cast_175")];
            tensor<fp16, [1, ?, 1024]> var_529_cast_fp16 = mul(x = x_45_cast_fp16, y = var_528_to_fp16)[name = tensor<string, []>("op_529_cast_fp16")];
            tensor<fp16, [1024]> blocks_4_att_norm_weight_to_fp16 = const()[name = tensor<string, []>("blocks_4_att_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(145864128)))];
            tensor<fp16, [1, ?, 1024]> input_69_cast_fp16 = mul(x = var_529_cast_fp16, y = blocks_4_att_norm_weight_to_fp16)[name = tensor<string, []>("input_69_cast_fp16")];
            tensor<fp16, [3072, 1024]> blocks_4_att_c_attn_weight_to_fp16 = const()[name = tensor<string, []>("blocks_4_att_c_attn_weight_to_fp16"), val = tensor<fp16, [3072, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(145866240)))];
            tensor<fp16, [1, ?, 3072]> linear_18_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = blocks_4_att_c_attn_weight_to_fp16, x = input_69_cast_fp16)[name = tensor<string, []>("linear_18_cast_fp16")];
            tensor<int32, [3]> tile_4 = const()[name = tensor<string, []>("tile_4"), val = tensor<int32, [3]>([1024, 1024, 1024])];
            tensor<int32, []> var_537_axis_0 = const()[name = tensor<string, []>("op_537_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, ?, 1024]> var_537_cast_fp16_0, tensor<fp16, [1, ?, 1024]> var_537_cast_fp16_1, tensor<fp16, [1, ?, 1024]> var_537_cast_fp16_2 = split(axis = var_537_axis_0, split_sizes = tile_4, x = linear_18_cast_fp16)[name = tensor<string, []>("op_537_cast_fp16")];
            tensor<int32, [4]> var_541 = const()[name = tensor<string, []>("op_541"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_542_cast_fp16 = reshape(shape = var_541, x = var_537_cast_fp16_0)[name = tensor<string, []>("op_542_cast_fp16")];
            tensor<int32, [4]> x_47_perm_0 = const()[name = tensor<string, []>("x_47_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_544 = const()[name = tensor<string, []>("op_544"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_545_cast_fp16 = reshape(shape = var_544, x = var_537_cast_fp16_1)[name = tensor<string, []>("op_545_cast_fp16")];
            tensor<int32, [4]> x_49_perm_0 = const()[name = tensor<string, []>("x_49_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_547 = const()[name = tensor<string, []>("op_547"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_548_cast_fp16 = reshape(shape = var_547, x = var_537_cast_fp16_2)[name = tensor<string, []>("op_548_cast_fp16")];
            tensor<int32, [4]> v_19_perm_0 = const()[name = tensor<string, []>("v_19_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> x1_17_begin_0 = const()[name = tensor<string, []>("x1_17_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_17_end_0 = const()[name = tensor<string, []>("x1_17_end_0"), val = tensor<int32, [4]>([1, 16, 0, 64])];
            tensor<bool, [4]> x1_17_end_mask_0 = const()[name = tensor<string, []>("x1_17_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> x1_17_stride_0 = const()[name = tensor<string, []>("x1_17_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 16, ?, 64]> x_47_cast_fp16 = transpose(perm = x_47_perm_0, x = var_542_cast_fp16)[name = tensor<string, []>("transpose_34")];
            tensor<fp16, [1, 16, ?, 32]> x1_17_cast_fp16 = slice_by_index(begin = x1_17_begin_0, end = x1_17_end_0, end_mask = x1_17_end_mask_0, stride = x1_17_stride_0, x = x_47_cast_fp16)[name = tensor<string, []>("x1_17_cast_fp16")];
            tensor<int32, [4]> x2_17_begin_0 = const()[name = tensor<string, []>("x2_17_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 1])];
            tensor<int32, [4]> x2_17_end_0 = const()[name = tensor<string, []>("x2_17_end_0"), val = tensor<int32, [4]>([1, 16, 0, 64])];
            tensor<bool, [4]> x2_17_end_mask_0 = const()[name = tensor<string, []>("x2_17_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> x2_17_stride_0 = const()[name = tensor<string, []>("x2_17_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 16, ?, 32]> x2_17_cast_fp16 = slice_by_index(begin = x2_17_begin_0, end = x2_17_end_0, end_mask = x2_17_end_mask_0, stride = x2_17_stride_0, x = x_47_cast_fp16)[name = tensor<string, []>("x2_17_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_552_cast_fp16 = mul(x = x1_17_cast_fp16, y = blocks_0_att_head_cos_to_fp16)[name = tensor<string, []>("op_552_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_553_cast_fp16 = mul(x = x2_17_cast_fp16, y = blocks_0_att_head_sin_to_fp16)[name = tensor<string, []>("op_553_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_554_cast_fp16 = sub(x = var_552_cast_fp16, y = var_553_cast_fp16)[name = tensor<string, []>("op_554_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_555_cast_fp16 = mul(x = x2_17_cast_fp16, y = blocks_0_att_head_cos_to_fp16)[name = tensor<string, []>("op_555_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_556_cast_fp16 = mul(x = x1_17_cast_fp16, y = blocks_0_att_head_sin_to_fp16)[name = tensor<string, []>("op_556_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_557_cast_fp16 = add(x = var_555_cast_fp16, y = var_556_cast_fp16)[name = tensor<string, []>("op_557_cast_fp16")];
            tensor<int32, []> out_17_axis_0 = const()[name = tensor<string, []>("out_17_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 16, ?, 32, 2]> out_17_cast_fp16 = stack(axis = out_17_axis_0, values = (var_554_cast_fp16, var_557_cast_fp16))[name = tensor<string, []>("out_17_cast_fp16")];
            tensor<int32, [4]> concat_12x = const()[name = tensor<string, []>("concat_12x"), val = tensor<int32, [4]>([1, 16, -1, 64])];
            tensor<fp16, [1, 16, ?, 64]> q_19_cast_fp16 = reshape(shape = concat_12x, x = out_17_cast_fp16)[name = tensor<string, []>("q_19_cast_fp16")];
            tensor<int32, [4]> x1_19_begin_0 = const()[name = tensor<string, []>("x1_19_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_19_end_0 = const()[name = tensor<string, []>("x1_19_end_0"), val = tensor<int32, [4]>([1, 16, 0, 64])];
            tensor<bool, [4]> x1_19_end_mask_0 = const()[name = tensor<string, []>("x1_19_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> x1_19_stride_0 = const()[name = tensor<string, []>("x1_19_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 16, ?, 64]> x_49_cast_fp16 = transpose(perm = x_49_perm_0, x = var_545_cast_fp16)[name = tensor<string, []>("transpose_33")];
            tensor<fp16, [1, 16, ?, 32]> x1_19_cast_fp16 = slice_by_index(begin = x1_19_begin_0, end = x1_19_end_0, end_mask = x1_19_end_mask_0, stride = x1_19_stride_0, x = x_49_cast_fp16)[name = tensor<string, []>("x1_19_cast_fp16")];
            tensor<int32, [4]> x2_19_begin_0 = const()[name = tensor<string, []>("x2_19_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 1])];
            tensor<int32, [4]> x2_19_end_0 = const()[name = tensor<string, []>("x2_19_end_0"), val = tensor<int32, [4]>([1, 16, 0, 64])];
            tensor<bool, [4]> x2_19_end_mask_0 = const()[name = tensor<string, []>("x2_19_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> x2_19_stride_0 = const()[name = tensor<string, []>("x2_19_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 16, ?, 32]> x2_19_cast_fp16 = slice_by_index(begin = x2_19_begin_0, end = x2_19_end_0, end_mask = x2_19_end_mask_0, stride = x2_19_stride_0, x = x_49_cast_fp16)[name = tensor<string, []>("x2_19_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_563_cast_fp16 = mul(x = x1_19_cast_fp16, y = blocks_0_att_head_cos_to_fp16)[name = tensor<string, []>("op_563_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_564_cast_fp16 = mul(x = x2_19_cast_fp16, y = blocks_0_att_head_sin_to_fp16)[name = tensor<string, []>("op_564_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_565_cast_fp16 = sub(x = var_563_cast_fp16, y = var_564_cast_fp16)[name = tensor<string, []>("op_565_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_566_cast_fp16 = mul(x = x2_19_cast_fp16, y = blocks_0_att_head_cos_to_fp16)[name = tensor<string, []>("op_566_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_567_cast_fp16 = mul(x = x1_19_cast_fp16, y = blocks_0_att_head_sin_to_fp16)[name = tensor<string, []>("op_567_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_568_cast_fp16 = add(x = var_566_cast_fp16, y = var_567_cast_fp16)[name = tensor<string, []>("op_568_cast_fp16")];
            tensor<int32, []> out_19_axis_0 = const()[name = tensor<string, []>("out_19_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 16, ?, 32, 2]> out_19_cast_fp16 = stack(axis = out_19_axis_0, values = (var_565_cast_fp16, var_568_cast_fp16))[name = tensor<string, []>("out_19_cast_fp16")];
            tensor<int32, [4]> concat_13x = const()[name = tensor<string, []>("concat_13x"), val = tensor<int32, [4]>([1, 16, -1, 64])];
            tensor<fp16, [1, 16, ?, 64]> k_19_cast_fp16 = reshape(shape = concat_13x, x = out_19_cast_fp16)[name = tensor<string, []>("k_19_cast_fp16")];
            tensor<fp16, []> mul_18_y_0_to_fp16 = const()[name = tensor<string, []>("mul_18_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 16, ?, 64]> mul_18_cast_fp16 = mul(x = q_19_cast_fp16, y = mul_18_y_0_to_fp16)[name = tensor<string, []>("mul_18_cast_fp16")];
            tensor<bool, []> matmul_4_transpose_y_0 = const()[name = tensor<string, []>("matmul_4_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_4_transpose_x_0 = const()[name = tensor<string, []>("matmul_4_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 16, ?, ?]> matmul_4_cast_fp16 = matmul(transpose_x = matmul_4_transpose_x_0, transpose_y = matmul_4_transpose_y_0, x = mul_18_cast_fp16, y = k_19_cast_fp16)[name = tensor<string, []>("matmul_4_cast_fp16")];
            tensor<string, []> matmul_4_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("matmul_4_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<int32, []> softmax_4_axis_0 = const()[name = tensor<string, []>("softmax_4_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp32, [1, 16, ?, ?]> matmul_4_cast_fp16_to_fp32 = cast(dtype = matmul_4_cast_fp16_to_fp32_dtype_0, x = matmul_4_cast_fp16)[name = tensor<string, []>("cast_174")];
            tensor<fp32, [1, 16, ?, ?]> softmax_4 = softmax(axis = softmax_4_axis_0, x = matmul_4_cast_fp16_to_fp32)[name = tensor<string, []>("softmax_4")];
            tensor<bool, []> y_9_transpose_x_0 = const()[name = tensor<string, []>("y_9_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> y_9_transpose_y_0 = const()[name = tensor<string, []>("y_9_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<string, []> softmax_4_to_fp16_dtype_0 = const()[name = tensor<string, []>("softmax_4_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 16, ?, ?]> softmax_4_to_fp16 = cast(dtype = softmax_4_to_fp16_dtype_0, x = softmax_4)[name = tensor<string, []>("cast_173")];
            tensor<fp16, [1, 16, ?, 64]> v_19_cast_fp16 = transpose(perm = v_19_perm_0, x = var_548_cast_fp16)[name = tensor<string, []>("transpose_32")];
            tensor<fp16, [1, 16, ?, 64]> y_9_cast_fp16 = matmul(transpose_x = y_9_transpose_x_0, transpose_y = y_9_transpose_y_0, x = softmax_4_to_fp16, y = v_19_cast_fp16)[name = tensor<string, []>("y_9_cast_fp16")];
            tensor<int32, [4]> var_573_perm_0 = const()[name = tensor<string, []>("op_573_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_574 = const()[name = tensor<string, []>("op_574"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, ?, 16, 64]> var_573_cast_fp16 = transpose(perm = var_573_perm_0, x = y_9_cast_fp16)[name = tensor<string, []>("transpose_31")];
            tensor<fp16, [1, ?, 1024]> input_71_cast_fp16 = reshape(shape = var_574, x = var_573_cast_fp16)[name = tensor<string, []>("input_71_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_4_att_c_proj_weight_to_fp16 = const()[name = tensor<string, []>("blocks_4_att_c_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(152157760)))];
            tensor<fp16, [1, ?, 1024]> linear_19_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = blocks_4_att_c_proj_weight_to_fp16, x = input_71_cast_fp16)[name = tensor<string, []>("linear_19_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_51_cast_fp16 = add(x = x_45_cast_fp16, y = linear_19_cast_fp16)[name = tensor<string, []>("x_51_cast_fp16")];
            tensor<string, []> x_51_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_51_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_513_promoted_1 = const()[name = tensor<string, []>("op_513_promoted_1"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, ?, 1024]> x_51_cast_fp16_to_fp32 = cast(dtype = x_51_cast_fp16_to_fp32_dtype_0, x = x_51_cast_fp16)[name = tensor<string, []>("cast_172")];
            tensor<fp32, [1, ?, 1024]> var_580 = pow(x = x_51_cast_fp16_to_fp32, y = var_513_promoted_1)[name = tensor<string, []>("op_580")];
            tensor<int32, [1]> norm_x_19_axes_0 = const()[name = tensor<string, []>("norm_x_19_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> norm_x_19_keep_dims_0 = const()[name = tensor<string, []>("norm_x_19_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, ?, 1]> norm_x_19 = reduce_mean(axes = norm_x_19_axes_0, keep_dims = norm_x_19_keep_dims_0, x = var_580)[name = tensor<string, []>("norm_x_19")];
            tensor<string, []> norm_x_19_to_fp16_dtype_0 = const()[name = tensor<string, []>("norm_x_19_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_583_to_fp16 = const()[name = tensor<string, []>("op_583_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> norm_x_19_to_fp16 = cast(dtype = norm_x_19_to_fp16_dtype_0, x = norm_x_19)[name = tensor<string, []>("cast_171")];
            tensor<fp16, [1, ?, 1]> var_584_cast_fp16 = add(x = norm_x_19_to_fp16, y = var_583_to_fp16)[name = tensor<string, []>("op_584_cast_fp16")];
            tensor<string, []> var_584_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_584_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_585_epsilon_0 = const()[name = tensor<string, []>("op_585_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, ?, 1]> var_584_cast_fp16_to_fp32 = cast(dtype = var_584_cast_fp16_to_fp32_dtype_0, x = var_584_cast_fp16)[name = tensor<string, []>("cast_170")];
            tensor<fp32, [1, ?, 1]> var_585 = rsqrt(epsilon = var_585_epsilon_0, x = var_584_cast_fp16_to_fp32)[name = tensor<string, []>("op_585")];
            tensor<string, []> var_585_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_585_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, ?, 1]> var_585_to_fp16 = cast(dtype = var_585_to_fp16_dtype_0, x = var_585)[name = tensor<string, []>("cast_169")];
            tensor<fp16, [1, ?, 1024]> var_586_cast_fp16 = mul(x = x_51_cast_fp16, y = var_585_to_fp16)[name = tensor<string, []>("op_586_cast_fp16")];
            tensor<fp16, [1024]> blocks_4_ffn_norm_weight_to_fp16 = const()[name = tensor<string, []>("blocks_4_ffn_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(154254976)))];
            tensor<fp16, [1, ?, 1024]> input_73_cast_fp16 = mul(x = var_586_cast_fp16, y = blocks_4_ffn_norm_weight_to_fp16)[name = tensor<string, []>("input_73_cast_fp16")];
            tensor<fp16, [4096, 1024]> blocks_4_mlp_fc1_weight_to_fp16 = const()[name = tensor<string, []>("blocks_4_mlp_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(154257088)))];
            tensor<fp16, [1, ?, 4096]> linear_20_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_4_mlp_fc1_weight_to_fp16, x = input_73_cast_fp16)[name = tensor<string, []>("linear_20_cast_fp16")];
            tensor<fp16, [1, ?, 4096]> input_77_cast_fp16 = silu(x = linear_20_cast_fp16)[name = tensor<string, []>("input_77_cast_fp16")];
            tensor<fp16, [1024, 4096]> blocks_4_mlp_fc2_weight_to_fp16 = const()[name = tensor<string, []>("blocks_4_mlp_fc2_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(162645760)))];
            tensor<fp16, [1, ?, 1024]> linear_21_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = blocks_4_mlp_fc2_weight_to_fp16, x = input_77_cast_fp16)[name = tensor<string, []>("linear_21_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_53_cast_fp16 = add(x = x_51_cast_fp16, y = linear_21_cast_fp16)[name = tensor<string, []>("x_53_cast_fp16")];
            tensor<string, []> x_53_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_53_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_605_promoted = const()[name = tensor<string, []>("op_605_promoted"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, ?, 1024]> x_53_cast_fp16_to_fp32 = cast(dtype = x_53_cast_fp16_to_fp32_dtype_0, x = x_53_cast_fp16)[name = tensor<string, []>("cast_168")];
            tensor<fp32, [1, ?, 1024]> var_615 = pow(x = x_53_cast_fp16_to_fp32, y = var_605_promoted)[name = tensor<string, []>("op_615")];
            tensor<int32, [1]> norm_x_21_axes_0 = const()[name = tensor<string, []>("norm_x_21_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> norm_x_21_keep_dims_0 = const()[name = tensor<string, []>("norm_x_21_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, ?, 1]> norm_x_21 = reduce_mean(axes = norm_x_21_axes_0, keep_dims = norm_x_21_keep_dims_0, x = var_615)[name = tensor<string, []>("norm_x_21")];
            tensor<string, []> norm_x_21_to_fp16_dtype_0 = const()[name = tensor<string, []>("norm_x_21_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_618_to_fp16 = const()[name = tensor<string, []>("op_618_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> norm_x_21_to_fp16 = cast(dtype = norm_x_21_to_fp16_dtype_0, x = norm_x_21)[name = tensor<string, []>("cast_167")];
            tensor<fp16, [1, ?, 1]> var_619_cast_fp16 = add(x = norm_x_21_to_fp16, y = var_618_to_fp16)[name = tensor<string, []>("op_619_cast_fp16")];
            tensor<string, []> var_619_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_619_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_620_epsilon_0 = const()[name = tensor<string, []>("op_620_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, ?, 1]> var_619_cast_fp16_to_fp32 = cast(dtype = var_619_cast_fp16_to_fp32_dtype_0, x = var_619_cast_fp16)[name = tensor<string, []>("cast_166")];
            tensor<fp32, [1, ?, 1]> var_620 = rsqrt(epsilon = var_620_epsilon_0, x = var_619_cast_fp16_to_fp32)[name = tensor<string, []>("op_620")];
            tensor<string, []> var_620_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_620_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, ?, 1]> var_620_to_fp16 = cast(dtype = var_620_to_fp16_dtype_0, x = var_620)[name = tensor<string, []>("cast_165")];
            tensor<fp16, [1, ?, 1024]> var_621_cast_fp16 = mul(x = x_53_cast_fp16, y = var_620_to_fp16)[name = tensor<string, []>("op_621_cast_fp16")];
            tensor<fp16, [1024]> blocks_5_att_norm_weight_to_fp16 = const()[name = tensor<string, []>("blocks_5_att_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(171034432)))];
            tensor<fp16, [1, ?, 1024]> input_79_cast_fp16 = mul(x = var_621_cast_fp16, y = blocks_5_att_norm_weight_to_fp16)[name = tensor<string, []>("input_79_cast_fp16")];
            tensor<fp16, [3072, 1024]> blocks_5_att_c_attn_weight_to_fp16 = const()[name = tensor<string, []>("blocks_5_att_c_attn_weight_to_fp16"), val = tensor<fp16, [3072, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(171036544)))];
            tensor<fp16, [1, ?, 3072]> linear_22_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = blocks_5_att_c_attn_weight_to_fp16, x = input_79_cast_fp16)[name = tensor<string, []>("linear_22_cast_fp16")];
            tensor<int32, [3]> tile_5 = const()[name = tensor<string, []>("tile_5"), val = tensor<int32, [3]>([1024, 1024, 1024])];
            tensor<int32, []> var_629_axis_0 = const()[name = tensor<string, []>("op_629_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, ?, 1024]> var_629_cast_fp16_0, tensor<fp16, [1, ?, 1024]> var_629_cast_fp16_1, tensor<fp16, [1, ?, 1024]> var_629_cast_fp16_2 = split(axis = var_629_axis_0, split_sizes = tile_5, x = linear_22_cast_fp16)[name = tensor<string, []>("op_629_cast_fp16")];
            tensor<int32, [4]> var_633 = const()[name = tensor<string, []>("op_633"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_634_cast_fp16 = reshape(shape = var_633, x = var_629_cast_fp16_0)[name = tensor<string, []>("op_634_cast_fp16")];
            tensor<int32, [4]> x_55_perm_0 = const()[name = tensor<string, []>("x_55_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_636 = const()[name = tensor<string, []>("op_636"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_637_cast_fp16 = reshape(shape = var_636, x = var_629_cast_fp16_1)[name = tensor<string, []>("op_637_cast_fp16")];
            tensor<int32, [4]> x_57_perm_0 = const()[name = tensor<string, []>("x_57_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_639 = const()[name = tensor<string, []>("op_639"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_640_cast_fp16 = reshape(shape = var_639, x = var_629_cast_fp16_2)[name = tensor<string, []>("op_640_cast_fp16")];
            tensor<int32, [4]> v_23_perm_0 = const()[name = tensor<string, []>("v_23_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> x1_21_begin_0 = const()[name = tensor<string, []>("x1_21_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_21_end_0 = const()[name = tensor<string, []>("x1_21_end_0"), val = tensor<int32, [4]>([1, 16, 0, 64])];
            tensor<bool, [4]> x1_21_end_mask_0 = const()[name = tensor<string, []>("x1_21_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> x1_21_stride_0 = const()[name = tensor<string, []>("x1_21_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 16, ?, 64]> x_55_cast_fp16 = transpose(perm = x_55_perm_0, x = var_634_cast_fp16)[name = tensor<string, []>("transpose_30")];
            tensor<fp16, [1, 16, ?, 32]> x1_21_cast_fp16 = slice_by_index(begin = x1_21_begin_0, end = x1_21_end_0, end_mask = x1_21_end_mask_0, stride = x1_21_stride_0, x = x_55_cast_fp16)[name = tensor<string, []>("x1_21_cast_fp16")];
            tensor<int32, [4]> x2_21_begin_0 = const()[name = tensor<string, []>("x2_21_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 1])];
            tensor<int32, [4]> x2_21_end_0 = const()[name = tensor<string, []>("x2_21_end_0"), val = tensor<int32, [4]>([1, 16, 0, 64])];
            tensor<bool, [4]> x2_21_end_mask_0 = const()[name = tensor<string, []>("x2_21_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> x2_21_stride_0 = const()[name = tensor<string, []>("x2_21_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 16, ?, 32]> x2_21_cast_fp16 = slice_by_index(begin = x2_21_begin_0, end = x2_21_end_0, end_mask = x2_21_end_mask_0, stride = x2_21_stride_0, x = x_55_cast_fp16)[name = tensor<string, []>("x2_21_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_644_cast_fp16 = mul(x = x1_21_cast_fp16, y = blocks_0_att_head_cos_to_fp16)[name = tensor<string, []>("op_644_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_645_cast_fp16 = mul(x = x2_21_cast_fp16, y = blocks_0_att_head_sin_to_fp16)[name = tensor<string, []>("op_645_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_646_cast_fp16 = sub(x = var_644_cast_fp16, y = var_645_cast_fp16)[name = tensor<string, []>("op_646_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_647_cast_fp16 = mul(x = x2_21_cast_fp16, y = blocks_0_att_head_cos_to_fp16)[name = tensor<string, []>("op_647_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_648_cast_fp16 = mul(x = x1_21_cast_fp16, y = blocks_0_att_head_sin_to_fp16)[name = tensor<string, []>("op_648_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_649_cast_fp16 = add(x = var_647_cast_fp16, y = var_648_cast_fp16)[name = tensor<string, []>("op_649_cast_fp16")];
            tensor<int32, []> out_21_axis_0 = const()[name = tensor<string, []>("out_21_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 16, ?, 32, 2]> out_21_cast_fp16 = stack(axis = out_21_axis_0, values = (var_646_cast_fp16, var_649_cast_fp16))[name = tensor<string, []>("out_21_cast_fp16")];
            tensor<int32, [4]> concat_14x = const()[name = tensor<string, []>("concat_14x"), val = tensor<int32, [4]>([1, 16, -1, 64])];
            tensor<fp16, [1, 16, ?, 64]> q_23_cast_fp16 = reshape(shape = concat_14x, x = out_21_cast_fp16)[name = tensor<string, []>("q_23_cast_fp16")];
            tensor<int32, [4]> x1_23_begin_0 = const()[name = tensor<string, []>("x1_23_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_23_end_0 = const()[name = tensor<string, []>("x1_23_end_0"), val = tensor<int32, [4]>([1, 16, 0, 64])];
            tensor<bool, [4]> x1_23_end_mask_0 = const()[name = tensor<string, []>("x1_23_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> x1_23_stride_0 = const()[name = tensor<string, []>("x1_23_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 16, ?, 64]> x_57_cast_fp16 = transpose(perm = x_57_perm_0, x = var_637_cast_fp16)[name = tensor<string, []>("transpose_29")];
            tensor<fp16, [1, 16, ?, 32]> x1_23_cast_fp16 = slice_by_index(begin = x1_23_begin_0, end = x1_23_end_0, end_mask = x1_23_end_mask_0, stride = x1_23_stride_0, x = x_57_cast_fp16)[name = tensor<string, []>("x1_23_cast_fp16")];
            tensor<int32, [4]> x2_23_begin_0 = const()[name = tensor<string, []>("x2_23_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 1])];
            tensor<int32, [4]> x2_23_end_0 = const()[name = tensor<string, []>("x2_23_end_0"), val = tensor<int32, [4]>([1, 16, 0, 64])];
            tensor<bool, [4]> x2_23_end_mask_0 = const()[name = tensor<string, []>("x2_23_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> x2_23_stride_0 = const()[name = tensor<string, []>("x2_23_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 16, ?, 32]> x2_23_cast_fp16 = slice_by_index(begin = x2_23_begin_0, end = x2_23_end_0, end_mask = x2_23_end_mask_0, stride = x2_23_stride_0, x = x_57_cast_fp16)[name = tensor<string, []>("x2_23_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_655_cast_fp16 = mul(x = x1_23_cast_fp16, y = blocks_0_att_head_cos_to_fp16)[name = tensor<string, []>("op_655_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_656_cast_fp16 = mul(x = x2_23_cast_fp16, y = blocks_0_att_head_sin_to_fp16)[name = tensor<string, []>("op_656_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_657_cast_fp16 = sub(x = var_655_cast_fp16, y = var_656_cast_fp16)[name = tensor<string, []>("op_657_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_658_cast_fp16 = mul(x = x2_23_cast_fp16, y = blocks_0_att_head_cos_to_fp16)[name = tensor<string, []>("op_658_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_659_cast_fp16 = mul(x = x1_23_cast_fp16, y = blocks_0_att_head_sin_to_fp16)[name = tensor<string, []>("op_659_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_660_cast_fp16 = add(x = var_658_cast_fp16, y = var_659_cast_fp16)[name = tensor<string, []>("op_660_cast_fp16")];
            tensor<int32, []> out_23_axis_0 = const()[name = tensor<string, []>("out_23_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 16, ?, 32, 2]> out_23_cast_fp16 = stack(axis = out_23_axis_0, values = (var_657_cast_fp16, var_660_cast_fp16))[name = tensor<string, []>("out_23_cast_fp16")];
            tensor<int32, [4]> concat_15x = const()[name = tensor<string, []>("concat_15x"), val = tensor<int32, [4]>([1, 16, -1, 64])];
            tensor<fp16, [1, 16, ?, 64]> k_23_cast_fp16 = reshape(shape = concat_15x, x = out_23_cast_fp16)[name = tensor<string, []>("k_23_cast_fp16")];
            tensor<fp16, []> mul_21_y_0_to_fp16 = const()[name = tensor<string, []>("mul_21_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 16, ?, 64]> mul_21_cast_fp16 = mul(x = q_23_cast_fp16, y = mul_21_y_0_to_fp16)[name = tensor<string, []>("mul_21_cast_fp16")];
            tensor<bool, []> matmul_5_transpose_y_0 = const()[name = tensor<string, []>("matmul_5_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_5_transpose_x_0 = const()[name = tensor<string, []>("matmul_5_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 16, ?, ?]> matmul_5_cast_fp16 = matmul(transpose_x = matmul_5_transpose_x_0, transpose_y = matmul_5_transpose_y_0, x = mul_21_cast_fp16, y = k_23_cast_fp16)[name = tensor<string, []>("matmul_5_cast_fp16")];
            tensor<string, []> matmul_5_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("matmul_5_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<int32, []> softmax_5_axis_0 = const()[name = tensor<string, []>("softmax_5_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp32, [1, 16, ?, ?]> matmul_5_cast_fp16_to_fp32 = cast(dtype = matmul_5_cast_fp16_to_fp32_dtype_0, x = matmul_5_cast_fp16)[name = tensor<string, []>("cast_164")];
            tensor<fp32, [1, 16, ?, ?]> softmax_5 = softmax(axis = softmax_5_axis_0, x = matmul_5_cast_fp16_to_fp32)[name = tensor<string, []>("softmax_5")];
            tensor<bool, []> y_11_transpose_x_0 = const()[name = tensor<string, []>("y_11_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> y_11_transpose_y_0 = const()[name = tensor<string, []>("y_11_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<string, []> softmax_5_to_fp16_dtype_0 = const()[name = tensor<string, []>("softmax_5_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 16, ?, ?]> softmax_5_to_fp16 = cast(dtype = softmax_5_to_fp16_dtype_0, x = softmax_5)[name = tensor<string, []>("cast_163")];
            tensor<fp16, [1, 16, ?, 64]> v_23_cast_fp16 = transpose(perm = v_23_perm_0, x = var_640_cast_fp16)[name = tensor<string, []>("transpose_28")];
            tensor<fp16, [1, 16, ?, 64]> y_11_cast_fp16 = matmul(transpose_x = y_11_transpose_x_0, transpose_y = y_11_transpose_y_0, x = softmax_5_to_fp16, y = v_23_cast_fp16)[name = tensor<string, []>("y_11_cast_fp16")];
            tensor<int32, [4]> var_665_perm_0 = const()[name = tensor<string, []>("op_665_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_666 = const()[name = tensor<string, []>("op_666"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, ?, 16, 64]> var_665_cast_fp16 = transpose(perm = var_665_perm_0, x = y_11_cast_fp16)[name = tensor<string, []>("transpose_27")];
            tensor<fp16, [1, ?, 1024]> input_81_cast_fp16 = reshape(shape = var_666, x = var_665_cast_fp16)[name = tensor<string, []>("input_81_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_5_att_c_proj_weight_to_fp16 = const()[name = tensor<string, []>("blocks_5_att_c_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(177328064)))];
            tensor<fp16, [1, ?, 1024]> linear_23_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = blocks_5_att_c_proj_weight_to_fp16, x = input_81_cast_fp16)[name = tensor<string, []>("linear_23_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_59_cast_fp16 = add(x = x_53_cast_fp16, y = linear_23_cast_fp16)[name = tensor<string, []>("x_59_cast_fp16")];
            tensor<string, []> x_59_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_59_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_605_promoted_1 = const()[name = tensor<string, []>("op_605_promoted_1"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, ?, 1024]> x_59_cast_fp16_to_fp32 = cast(dtype = x_59_cast_fp16_to_fp32_dtype_0, x = x_59_cast_fp16)[name = tensor<string, []>("cast_162")];
            tensor<fp32, [1, ?, 1024]> var_672 = pow(x = x_59_cast_fp16_to_fp32, y = var_605_promoted_1)[name = tensor<string, []>("op_672")];
            tensor<int32, [1]> norm_x_23_axes_0 = const()[name = tensor<string, []>("norm_x_23_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> norm_x_23_keep_dims_0 = const()[name = tensor<string, []>("norm_x_23_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, ?, 1]> norm_x_23 = reduce_mean(axes = norm_x_23_axes_0, keep_dims = norm_x_23_keep_dims_0, x = var_672)[name = tensor<string, []>("norm_x_23")];
            tensor<string, []> norm_x_23_to_fp16_dtype_0 = const()[name = tensor<string, []>("norm_x_23_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_675_to_fp16 = const()[name = tensor<string, []>("op_675_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> norm_x_23_to_fp16 = cast(dtype = norm_x_23_to_fp16_dtype_0, x = norm_x_23)[name = tensor<string, []>("cast_161")];
            tensor<fp16, [1, ?, 1]> var_676_cast_fp16 = add(x = norm_x_23_to_fp16, y = var_675_to_fp16)[name = tensor<string, []>("op_676_cast_fp16")];
            tensor<string, []> var_676_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_676_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_677_epsilon_0 = const()[name = tensor<string, []>("op_677_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, ?, 1]> var_676_cast_fp16_to_fp32 = cast(dtype = var_676_cast_fp16_to_fp32_dtype_0, x = var_676_cast_fp16)[name = tensor<string, []>("cast_160")];
            tensor<fp32, [1, ?, 1]> var_677 = rsqrt(epsilon = var_677_epsilon_0, x = var_676_cast_fp16_to_fp32)[name = tensor<string, []>("op_677")];
            tensor<string, []> var_677_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_677_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, ?, 1]> var_677_to_fp16 = cast(dtype = var_677_to_fp16_dtype_0, x = var_677)[name = tensor<string, []>("cast_159")];
            tensor<fp16, [1, ?, 1024]> var_678_cast_fp16 = mul(x = x_59_cast_fp16, y = var_677_to_fp16)[name = tensor<string, []>("op_678_cast_fp16")];
            tensor<fp16, [1024]> blocks_5_ffn_norm_weight_to_fp16 = const()[name = tensor<string, []>("blocks_5_ffn_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(179425280)))];
            tensor<fp16, [1, ?, 1024]> input_83_cast_fp16 = mul(x = var_678_cast_fp16, y = blocks_5_ffn_norm_weight_to_fp16)[name = tensor<string, []>("input_83_cast_fp16")];
            tensor<fp16, [4096, 1024]> blocks_5_mlp_fc1_weight_to_fp16 = const()[name = tensor<string, []>("blocks_5_mlp_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(179427392)))];
            tensor<fp16, [1, ?, 4096]> linear_24_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_5_mlp_fc1_weight_to_fp16, x = input_83_cast_fp16)[name = tensor<string, []>("linear_24_cast_fp16")];
            tensor<fp16, [1, ?, 4096]> input_87_cast_fp16 = silu(x = linear_24_cast_fp16)[name = tensor<string, []>("input_87_cast_fp16")];
            tensor<fp16, [1024, 4096]> blocks_5_mlp_fc2_weight_to_fp16 = const()[name = tensor<string, []>("blocks_5_mlp_fc2_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(187816064)))];
            tensor<fp16, [1, ?, 1024]> linear_25_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = blocks_5_mlp_fc2_weight_to_fp16, x = input_87_cast_fp16)[name = tensor<string, []>("linear_25_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_61_cast_fp16 = add(x = x_59_cast_fp16, y = linear_25_cast_fp16)[name = tensor<string, []>("x_61_cast_fp16")];
            tensor<string, []> x_61_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_61_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_697_promoted = const()[name = tensor<string, []>("op_697_promoted"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, ?, 1024]> x_61_cast_fp16_to_fp32 = cast(dtype = x_61_cast_fp16_to_fp32_dtype_0, x = x_61_cast_fp16)[name = tensor<string, []>("cast_158")];
            tensor<fp32, [1, ?, 1024]> var_707 = pow(x = x_61_cast_fp16_to_fp32, y = var_697_promoted)[name = tensor<string, []>("op_707")];
            tensor<int32, [1]> norm_x_25_axes_0 = const()[name = tensor<string, []>("norm_x_25_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> norm_x_25_keep_dims_0 = const()[name = tensor<string, []>("norm_x_25_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, ?, 1]> norm_x_25 = reduce_mean(axes = norm_x_25_axes_0, keep_dims = norm_x_25_keep_dims_0, x = var_707)[name = tensor<string, []>("norm_x_25")];
            tensor<string, []> norm_x_25_to_fp16_dtype_0 = const()[name = tensor<string, []>("norm_x_25_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_710_to_fp16 = const()[name = tensor<string, []>("op_710_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> norm_x_25_to_fp16 = cast(dtype = norm_x_25_to_fp16_dtype_0, x = norm_x_25)[name = tensor<string, []>("cast_157")];
            tensor<fp16, [1, ?, 1]> var_711_cast_fp16 = add(x = norm_x_25_to_fp16, y = var_710_to_fp16)[name = tensor<string, []>("op_711_cast_fp16")];
            tensor<string, []> var_711_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_711_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_712_epsilon_0 = const()[name = tensor<string, []>("op_712_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, ?, 1]> var_711_cast_fp16_to_fp32 = cast(dtype = var_711_cast_fp16_to_fp32_dtype_0, x = var_711_cast_fp16)[name = tensor<string, []>("cast_156")];
            tensor<fp32, [1, ?, 1]> var_712 = rsqrt(epsilon = var_712_epsilon_0, x = var_711_cast_fp16_to_fp32)[name = tensor<string, []>("op_712")];
            tensor<string, []> var_712_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_712_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, ?, 1]> var_712_to_fp16 = cast(dtype = var_712_to_fp16_dtype_0, x = var_712)[name = tensor<string, []>("cast_155")];
            tensor<fp16, [1, ?, 1024]> var_713_cast_fp16 = mul(x = x_61_cast_fp16, y = var_712_to_fp16)[name = tensor<string, []>("op_713_cast_fp16")];
            tensor<fp16, [1024]> blocks_6_att_norm_weight_to_fp16 = const()[name = tensor<string, []>("blocks_6_att_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(196204736)))];
            tensor<fp16, [1, ?, 1024]> input_89_cast_fp16 = mul(x = var_713_cast_fp16, y = blocks_6_att_norm_weight_to_fp16)[name = tensor<string, []>("input_89_cast_fp16")];
            tensor<fp16, [3072, 1024]> blocks_6_att_c_attn_weight_to_fp16 = const()[name = tensor<string, []>("blocks_6_att_c_attn_weight_to_fp16"), val = tensor<fp16, [3072, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(196206848)))];
            tensor<fp16, [1, ?, 3072]> linear_26_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = blocks_6_att_c_attn_weight_to_fp16, x = input_89_cast_fp16)[name = tensor<string, []>("linear_26_cast_fp16")];
            tensor<int32, [3]> tile_6 = const()[name = tensor<string, []>("tile_6"), val = tensor<int32, [3]>([1024, 1024, 1024])];
            tensor<int32, []> var_721_axis_0 = const()[name = tensor<string, []>("op_721_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, ?, 1024]> var_721_cast_fp16_0, tensor<fp16, [1, ?, 1024]> var_721_cast_fp16_1, tensor<fp16, [1, ?, 1024]> var_721_cast_fp16_2 = split(axis = var_721_axis_0, split_sizes = tile_6, x = linear_26_cast_fp16)[name = tensor<string, []>("op_721_cast_fp16")];
            tensor<int32, [4]> var_725 = const()[name = tensor<string, []>("op_725"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_726_cast_fp16 = reshape(shape = var_725, x = var_721_cast_fp16_0)[name = tensor<string, []>("op_726_cast_fp16")];
            tensor<int32, [4]> x_63_perm_0 = const()[name = tensor<string, []>("x_63_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_728 = const()[name = tensor<string, []>("op_728"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_729_cast_fp16 = reshape(shape = var_728, x = var_721_cast_fp16_1)[name = tensor<string, []>("op_729_cast_fp16")];
            tensor<int32, [4]> x_65_perm_0 = const()[name = tensor<string, []>("x_65_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_731 = const()[name = tensor<string, []>("op_731"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_732_cast_fp16 = reshape(shape = var_731, x = var_721_cast_fp16_2)[name = tensor<string, []>("op_732_cast_fp16")];
            tensor<int32, [4]> v_27_perm_0 = const()[name = tensor<string, []>("v_27_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> x1_25_begin_0 = const()[name = tensor<string, []>("x1_25_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_25_end_0 = const()[name = tensor<string, []>("x1_25_end_0"), val = tensor<int32, [4]>([1, 16, 0, 64])];
            tensor<bool, [4]> x1_25_end_mask_0 = const()[name = tensor<string, []>("x1_25_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> x1_25_stride_0 = const()[name = tensor<string, []>("x1_25_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 16, ?, 64]> x_63_cast_fp16 = transpose(perm = x_63_perm_0, x = var_726_cast_fp16)[name = tensor<string, []>("transpose_26")];
            tensor<fp16, [1, 16, ?, 32]> x1_25_cast_fp16 = slice_by_index(begin = x1_25_begin_0, end = x1_25_end_0, end_mask = x1_25_end_mask_0, stride = x1_25_stride_0, x = x_63_cast_fp16)[name = tensor<string, []>("x1_25_cast_fp16")];
            tensor<int32, [4]> x2_25_begin_0 = const()[name = tensor<string, []>("x2_25_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 1])];
            tensor<int32, [4]> x2_25_end_0 = const()[name = tensor<string, []>("x2_25_end_0"), val = tensor<int32, [4]>([1, 16, 0, 64])];
            tensor<bool, [4]> x2_25_end_mask_0 = const()[name = tensor<string, []>("x2_25_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> x2_25_stride_0 = const()[name = tensor<string, []>("x2_25_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 16, ?, 32]> x2_25_cast_fp16 = slice_by_index(begin = x2_25_begin_0, end = x2_25_end_0, end_mask = x2_25_end_mask_0, stride = x2_25_stride_0, x = x_63_cast_fp16)[name = tensor<string, []>("x2_25_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_736_cast_fp16 = mul(x = x1_25_cast_fp16, y = blocks_0_att_head_cos_to_fp16)[name = tensor<string, []>("op_736_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_737_cast_fp16 = mul(x = x2_25_cast_fp16, y = blocks_0_att_head_sin_to_fp16)[name = tensor<string, []>("op_737_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_738_cast_fp16 = sub(x = var_736_cast_fp16, y = var_737_cast_fp16)[name = tensor<string, []>("op_738_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_739_cast_fp16 = mul(x = x2_25_cast_fp16, y = blocks_0_att_head_cos_to_fp16)[name = tensor<string, []>("op_739_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_740_cast_fp16 = mul(x = x1_25_cast_fp16, y = blocks_0_att_head_sin_to_fp16)[name = tensor<string, []>("op_740_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_741_cast_fp16 = add(x = var_739_cast_fp16, y = var_740_cast_fp16)[name = tensor<string, []>("op_741_cast_fp16")];
            tensor<int32, []> out_25_axis_0 = const()[name = tensor<string, []>("out_25_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 16, ?, 32, 2]> out_25_cast_fp16 = stack(axis = out_25_axis_0, values = (var_738_cast_fp16, var_741_cast_fp16))[name = tensor<string, []>("out_25_cast_fp16")];
            tensor<int32, [4]> concat_16x = const()[name = tensor<string, []>("concat_16x"), val = tensor<int32, [4]>([1, 16, -1, 64])];
            tensor<fp16, [1, 16, ?, 64]> q_27_cast_fp16 = reshape(shape = concat_16x, x = out_25_cast_fp16)[name = tensor<string, []>("q_27_cast_fp16")];
            tensor<int32, [4]> x1_27_begin_0 = const()[name = tensor<string, []>("x1_27_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_27_end_0 = const()[name = tensor<string, []>("x1_27_end_0"), val = tensor<int32, [4]>([1, 16, 0, 64])];
            tensor<bool, [4]> x1_27_end_mask_0 = const()[name = tensor<string, []>("x1_27_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> x1_27_stride_0 = const()[name = tensor<string, []>("x1_27_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 16, ?, 64]> x_65_cast_fp16 = transpose(perm = x_65_perm_0, x = var_729_cast_fp16)[name = tensor<string, []>("transpose_25")];
            tensor<fp16, [1, 16, ?, 32]> x1_27_cast_fp16 = slice_by_index(begin = x1_27_begin_0, end = x1_27_end_0, end_mask = x1_27_end_mask_0, stride = x1_27_stride_0, x = x_65_cast_fp16)[name = tensor<string, []>("x1_27_cast_fp16")];
            tensor<int32, [4]> x2_27_begin_0 = const()[name = tensor<string, []>("x2_27_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 1])];
            tensor<int32, [4]> x2_27_end_0 = const()[name = tensor<string, []>("x2_27_end_0"), val = tensor<int32, [4]>([1, 16, 0, 64])];
            tensor<bool, [4]> x2_27_end_mask_0 = const()[name = tensor<string, []>("x2_27_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> x2_27_stride_0 = const()[name = tensor<string, []>("x2_27_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 16, ?, 32]> x2_27_cast_fp16 = slice_by_index(begin = x2_27_begin_0, end = x2_27_end_0, end_mask = x2_27_end_mask_0, stride = x2_27_stride_0, x = x_65_cast_fp16)[name = tensor<string, []>("x2_27_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_747_cast_fp16 = mul(x = x1_27_cast_fp16, y = blocks_0_att_head_cos_to_fp16)[name = tensor<string, []>("op_747_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_748_cast_fp16 = mul(x = x2_27_cast_fp16, y = blocks_0_att_head_sin_to_fp16)[name = tensor<string, []>("op_748_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_749_cast_fp16 = sub(x = var_747_cast_fp16, y = var_748_cast_fp16)[name = tensor<string, []>("op_749_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_750_cast_fp16 = mul(x = x2_27_cast_fp16, y = blocks_0_att_head_cos_to_fp16)[name = tensor<string, []>("op_750_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_751_cast_fp16 = mul(x = x1_27_cast_fp16, y = blocks_0_att_head_sin_to_fp16)[name = tensor<string, []>("op_751_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_752_cast_fp16 = add(x = var_750_cast_fp16, y = var_751_cast_fp16)[name = tensor<string, []>("op_752_cast_fp16")];
            tensor<int32, []> out_27_axis_0 = const()[name = tensor<string, []>("out_27_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 16, ?, 32, 2]> out_27_cast_fp16 = stack(axis = out_27_axis_0, values = (var_749_cast_fp16, var_752_cast_fp16))[name = tensor<string, []>("out_27_cast_fp16")];
            tensor<int32, [4]> concat_17x = const()[name = tensor<string, []>("concat_17x"), val = tensor<int32, [4]>([1, 16, -1, 64])];
            tensor<fp16, [1, 16, ?, 64]> k_27_cast_fp16 = reshape(shape = concat_17x, x = out_27_cast_fp16)[name = tensor<string, []>("k_27_cast_fp16")];
            tensor<fp16, []> mul_24_y_0_to_fp16 = const()[name = tensor<string, []>("mul_24_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 16, ?, 64]> mul_24_cast_fp16 = mul(x = q_27_cast_fp16, y = mul_24_y_0_to_fp16)[name = tensor<string, []>("mul_24_cast_fp16")];
            tensor<bool, []> matmul_6_transpose_y_0 = const()[name = tensor<string, []>("matmul_6_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_6_transpose_x_0 = const()[name = tensor<string, []>("matmul_6_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 16, ?, ?]> matmul_6_cast_fp16 = matmul(transpose_x = matmul_6_transpose_x_0, transpose_y = matmul_6_transpose_y_0, x = mul_24_cast_fp16, y = k_27_cast_fp16)[name = tensor<string, []>("matmul_6_cast_fp16")];
            tensor<string, []> matmul_6_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("matmul_6_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<int32, []> softmax_6_axis_0 = const()[name = tensor<string, []>("softmax_6_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp32, [1, 16, ?, ?]> matmul_6_cast_fp16_to_fp32 = cast(dtype = matmul_6_cast_fp16_to_fp32_dtype_0, x = matmul_6_cast_fp16)[name = tensor<string, []>("cast_154")];
            tensor<fp32, [1, 16, ?, ?]> softmax_6 = softmax(axis = softmax_6_axis_0, x = matmul_6_cast_fp16_to_fp32)[name = tensor<string, []>("softmax_6")];
            tensor<bool, []> y_13_transpose_x_0 = const()[name = tensor<string, []>("y_13_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> y_13_transpose_y_0 = const()[name = tensor<string, []>("y_13_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<string, []> softmax_6_to_fp16_dtype_0 = const()[name = tensor<string, []>("softmax_6_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 16, ?, ?]> softmax_6_to_fp16 = cast(dtype = softmax_6_to_fp16_dtype_0, x = softmax_6)[name = tensor<string, []>("cast_153")];
            tensor<fp16, [1, 16, ?, 64]> v_27_cast_fp16 = transpose(perm = v_27_perm_0, x = var_732_cast_fp16)[name = tensor<string, []>("transpose_24")];
            tensor<fp16, [1, 16, ?, 64]> y_13_cast_fp16 = matmul(transpose_x = y_13_transpose_x_0, transpose_y = y_13_transpose_y_0, x = softmax_6_to_fp16, y = v_27_cast_fp16)[name = tensor<string, []>("y_13_cast_fp16")];
            tensor<int32, [4]> var_757_perm_0 = const()[name = tensor<string, []>("op_757_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_758 = const()[name = tensor<string, []>("op_758"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, ?, 16, 64]> var_757_cast_fp16 = transpose(perm = var_757_perm_0, x = y_13_cast_fp16)[name = tensor<string, []>("transpose_23")];
            tensor<fp16, [1, ?, 1024]> input_91_cast_fp16 = reshape(shape = var_758, x = var_757_cast_fp16)[name = tensor<string, []>("input_91_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_6_att_c_proj_weight_to_fp16 = const()[name = tensor<string, []>("blocks_6_att_c_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(202498368)))];
            tensor<fp16, [1, ?, 1024]> linear_27_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = blocks_6_att_c_proj_weight_to_fp16, x = input_91_cast_fp16)[name = tensor<string, []>("linear_27_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_67_cast_fp16 = add(x = x_61_cast_fp16, y = linear_27_cast_fp16)[name = tensor<string, []>("x_67_cast_fp16")];
            tensor<string, []> x_67_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_67_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_697_promoted_1 = const()[name = tensor<string, []>("op_697_promoted_1"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, ?, 1024]> x_67_cast_fp16_to_fp32 = cast(dtype = x_67_cast_fp16_to_fp32_dtype_0, x = x_67_cast_fp16)[name = tensor<string, []>("cast_152")];
            tensor<fp32, [1, ?, 1024]> var_764 = pow(x = x_67_cast_fp16_to_fp32, y = var_697_promoted_1)[name = tensor<string, []>("op_764")];
            tensor<int32, [1]> norm_x_27_axes_0 = const()[name = tensor<string, []>("norm_x_27_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> norm_x_27_keep_dims_0 = const()[name = tensor<string, []>("norm_x_27_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, ?, 1]> norm_x_27 = reduce_mean(axes = norm_x_27_axes_0, keep_dims = norm_x_27_keep_dims_0, x = var_764)[name = tensor<string, []>("norm_x_27")];
            tensor<string, []> norm_x_27_to_fp16_dtype_0 = const()[name = tensor<string, []>("norm_x_27_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_767_to_fp16 = const()[name = tensor<string, []>("op_767_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> norm_x_27_to_fp16 = cast(dtype = norm_x_27_to_fp16_dtype_0, x = norm_x_27)[name = tensor<string, []>("cast_151")];
            tensor<fp16, [1, ?, 1]> var_768_cast_fp16 = add(x = norm_x_27_to_fp16, y = var_767_to_fp16)[name = tensor<string, []>("op_768_cast_fp16")];
            tensor<string, []> var_768_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_768_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_769_epsilon_0 = const()[name = tensor<string, []>("op_769_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, ?, 1]> var_768_cast_fp16_to_fp32 = cast(dtype = var_768_cast_fp16_to_fp32_dtype_0, x = var_768_cast_fp16)[name = tensor<string, []>("cast_150")];
            tensor<fp32, [1, ?, 1]> var_769 = rsqrt(epsilon = var_769_epsilon_0, x = var_768_cast_fp16_to_fp32)[name = tensor<string, []>("op_769")];
            tensor<string, []> var_769_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_769_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, ?, 1]> var_769_to_fp16 = cast(dtype = var_769_to_fp16_dtype_0, x = var_769)[name = tensor<string, []>("cast_149")];
            tensor<fp16, [1, ?, 1024]> var_770_cast_fp16 = mul(x = x_67_cast_fp16, y = var_769_to_fp16)[name = tensor<string, []>("op_770_cast_fp16")];
            tensor<fp16, [1024]> blocks_6_ffn_norm_weight_to_fp16 = const()[name = tensor<string, []>("blocks_6_ffn_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(204595584)))];
            tensor<fp16, [1, ?, 1024]> input_93_cast_fp16 = mul(x = var_770_cast_fp16, y = blocks_6_ffn_norm_weight_to_fp16)[name = tensor<string, []>("input_93_cast_fp16")];
            tensor<fp16, [4096, 1024]> blocks_6_mlp_fc1_weight_to_fp16 = const()[name = tensor<string, []>("blocks_6_mlp_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(204597696)))];
            tensor<fp16, [1, ?, 4096]> linear_28_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_6_mlp_fc1_weight_to_fp16, x = input_93_cast_fp16)[name = tensor<string, []>("linear_28_cast_fp16")];
            tensor<fp16, [1, ?, 4096]> input_97_cast_fp16 = silu(x = linear_28_cast_fp16)[name = tensor<string, []>("input_97_cast_fp16")];
            tensor<fp16, [1024, 4096]> blocks_6_mlp_fc2_weight_to_fp16 = const()[name = tensor<string, []>("blocks_6_mlp_fc2_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(212986368)))];
            tensor<fp16, [1, ?, 1024]> linear_29_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = blocks_6_mlp_fc2_weight_to_fp16, x = input_97_cast_fp16)[name = tensor<string, []>("linear_29_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_69_cast_fp16 = add(x = x_67_cast_fp16, y = linear_29_cast_fp16)[name = tensor<string, []>("x_69_cast_fp16")];
            tensor<string, []> x_69_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_69_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_789_promoted = const()[name = tensor<string, []>("op_789_promoted"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, ?, 1024]> x_69_cast_fp16_to_fp32 = cast(dtype = x_69_cast_fp16_to_fp32_dtype_0, x = x_69_cast_fp16)[name = tensor<string, []>("cast_148")];
            tensor<fp32, [1, ?, 1024]> var_799 = pow(x = x_69_cast_fp16_to_fp32, y = var_789_promoted)[name = tensor<string, []>("op_799")];
            tensor<int32, [1]> norm_x_29_axes_0 = const()[name = tensor<string, []>("norm_x_29_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> norm_x_29_keep_dims_0 = const()[name = tensor<string, []>("norm_x_29_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, ?, 1]> norm_x_29 = reduce_mean(axes = norm_x_29_axes_0, keep_dims = norm_x_29_keep_dims_0, x = var_799)[name = tensor<string, []>("norm_x_29")];
            tensor<string, []> norm_x_29_to_fp16_dtype_0 = const()[name = tensor<string, []>("norm_x_29_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_802_to_fp16 = const()[name = tensor<string, []>("op_802_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> norm_x_29_to_fp16 = cast(dtype = norm_x_29_to_fp16_dtype_0, x = norm_x_29)[name = tensor<string, []>("cast_147")];
            tensor<fp16, [1, ?, 1]> var_803_cast_fp16 = add(x = norm_x_29_to_fp16, y = var_802_to_fp16)[name = tensor<string, []>("op_803_cast_fp16")];
            tensor<string, []> var_803_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_803_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_804_epsilon_0 = const()[name = tensor<string, []>("op_804_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, ?, 1]> var_803_cast_fp16_to_fp32 = cast(dtype = var_803_cast_fp16_to_fp32_dtype_0, x = var_803_cast_fp16)[name = tensor<string, []>("cast_146")];
            tensor<fp32, [1, ?, 1]> var_804 = rsqrt(epsilon = var_804_epsilon_0, x = var_803_cast_fp16_to_fp32)[name = tensor<string, []>("op_804")];
            tensor<string, []> var_804_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_804_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, ?, 1]> var_804_to_fp16 = cast(dtype = var_804_to_fp16_dtype_0, x = var_804)[name = tensor<string, []>("cast_145")];
            tensor<fp16, [1, ?, 1024]> var_805_cast_fp16 = mul(x = x_69_cast_fp16, y = var_804_to_fp16)[name = tensor<string, []>("op_805_cast_fp16")];
            tensor<fp16, [1024]> blocks_7_att_norm_weight_to_fp16 = const()[name = tensor<string, []>("blocks_7_att_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(221375040)))];
            tensor<fp16, [1, ?, 1024]> input_99_cast_fp16 = mul(x = var_805_cast_fp16, y = blocks_7_att_norm_weight_to_fp16)[name = tensor<string, []>("input_99_cast_fp16")];
            tensor<fp16, [3072, 1024]> blocks_7_att_c_attn_weight_to_fp16 = const()[name = tensor<string, []>("blocks_7_att_c_attn_weight_to_fp16"), val = tensor<fp16, [3072, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(221377152)))];
            tensor<fp16, [1, ?, 3072]> linear_30_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = blocks_7_att_c_attn_weight_to_fp16, x = input_99_cast_fp16)[name = tensor<string, []>("linear_30_cast_fp16")];
            tensor<int32, [3]> tile_7 = const()[name = tensor<string, []>("tile_7"), val = tensor<int32, [3]>([1024, 1024, 1024])];
            tensor<int32, []> var_813_axis_0 = const()[name = tensor<string, []>("op_813_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, ?, 1024]> var_813_cast_fp16_0, tensor<fp16, [1, ?, 1024]> var_813_cast_fp16_1, tensor<fp16, [1, ?, 1024]> var_813_cast_fp16_2 = split(axis = var_813_axis_0, split_sizes = tile_7, x = linear_30_cast_fp16)[name = tensor<string, []>("op_813_cast_fp16")];
            tensor<int32, [4]> var_817 = const()[name = tensor<string, []>("op_817"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_818_cast_fp16 = reshape(shape = var_817, x = var_813_cast_fp16_0)[name = tensor<string, []>("op_818_cast_fp16")];
            tensor<int32, [4]> x_71_perm_0 = const()[name = tensor<string, []>("x_71_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_820 = const()[name = tensor<string, []>("op_820"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_821_cast_fp16 = reshape(shape = var_820, x = var_813_cast_fp16_1)[name = tensor<string, []>("op_821_cast_fp16")];
            tensor<int32, [4]> x_73_perm_0 = const()[name = tensor<string, []>("x_73_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_823 = const()[name = tensor<string, []>("op_823"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_824_cast_fp16 = reshape(shape = var_823, x = var_813_cast_fp16_2)[name = tensor<string, []>("op_824_cast_fp16")];
            tensor<int32, [4]> v_31_perm_0 = const()[name = tensor<string, []>("v_31_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> x1_29_begin_0 = const()[name = tensor<string, []>("x1_29_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_29_end_0 = const()[name = tensor<string, []>("x1_29_end_0"), val = tensor<int32, [4]>([1, 16, 0, 64])];
            tensor<bool, [4]> x1_29_end_mask_0 = const()[name = tensor<string, []>("x1_29_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> x1_29_stride_0 = const()[name = tensor<string, []>("x1_29_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 16, ?, 64]> x_71_cast_fp16 = transpose(perm = x_71_perm_0, x = var_818_cast_fp16)[name = tensor<string, []>("transpose_22")];
            tensor<fp16, [1, 16, ?, 32]> x1_29_cast_fp16 = slice_by_index(begin = x1_29_begin_0, end = x1_29_end_0, end_mask = x1_29_end_mask_0, stride = x1_29_stride_0, x = x_71_cast_fp16)[name = tensor<string, []>("x1_29_cast_fp16")];
            tensor<int32, [4]> x2_29_begin_0 = const()[name = tensor<string, []>("x2_29_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 1])];
            tensor<int32, [4]> x2_29_end_0 = const()[name = tensor<string, []>("x2_29_end_0"), val = tensor<int32, [4]>([1, 16, 0, 64])];
            tensor<bool, [4]> x2_29_end_mask_0 = const()[name = tensor<string, []>("x2_29_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> x2_29_stride_0 = const()[name = tensor<string, []>("x2_29_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 16, ?, 32]> x2_29_cast_fp16 = slice_by_index(begin = x2_29_begin_0, end = x2_29_end_0, end_mask = x2_29_end_mask_0, stride = x2_29_stride_0, x = x_71_cast_fp16)[name = tensor<string, []>("x2_29_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_828_cast_fp16 = mul(x = x1_29_cast_fp16, y = blocks_0_att_head_cos_to_fp16)[name = tensor<string, []>("op_828_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_829_cast_fp16 = mul(x = x2_29_cast_fp16, y = blocks_0_att_head_sin_to_fp16)[name = tensor<string, []>("op_829_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_830_cast_fp16 = sub(x = var_828_cast_fp16, y = var_829_cast_fp16)[name = tensor<string, []>("op_830_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_831_cast_fp16 = mul(x = x2_29_cast_fp16, y = blocks_0_att_head_cos_to_fp16)[name = tensor<string, []>("op_831_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_832_cast_fp16 = mul(x = x1_29_cast_fp16, y = blocks_0_att_head_sin_to_fp16)[name = tensor<string, []>("op_832_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_833_cast_fp16 = add(x = var_831_cast_fp16, y = var_832_cast_fp16)[name = tensor<string, []>("op_833_cast_fp16")];
            tensor<int32, []> out_29_axis_0 = const()[name = tensor<string, []>("out_29_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 16, ?, 32, 2]> out_29_cast_fp16 = stack(axis = out_29_axis_0, values = (var_830_cast_fp16, var_833_cast_fp16))[name = tensor<string, []>("out_29_cast_fp16")];
            tensor<int32, [4]> concat_18x = const()[name = tensor<string, []>("concat_18x"), val = tensor<int32, [4]>([1, 16, -1, 64])];
            tensor<fp16, [1, 16, ?, 64]> q_31_cast_fp16 = reshape(shape = concat_18x, x = out_29_cast_fp16)[name = tensor<string, []>("q_31_cast_fp16")];
            tensor<int32, [4]> x1_31_begin_0 = const()[name = tensor<string, []>("x1_31_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_31_end_0 = const()[name = tensor<string, []>("x1_31_end_0"), val = tensor<int32, [4]>([1, 16, 0, 64])];
            tensor<bool, [4]> x1_31_end_mask_0 = const()[name = tensor<string, []>("x1_31_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> x1_31_stride_0 = const()[name = tensor<string, []>("x1_31_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 16, ?, 64]> x_73_cast_fp16 = transpose(perm = x_73_perm_0, x = var_821_cast_fp16)[name = tensor<string, []>("transpose_21")];
            tensor<fp16, [1, 16, ?, 32]> x1_31_cast_fp16 = slice_by_index(begin = x1_31_begin_0, end = x1_31_end_0, end_mask = x1_31_end_mask_0, stride = x1_31_stride_0, x = x_73_cast_fp16)[name = tensor<string, []>("x1_31_cast_fp16")];
            tensor<int32, [4]> x2_31_begin_0 = const()[name = tensor<string, []>("x2_31_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 1])];
            tensor<int32, [4]> x2_31_end_0 = const()[name = tensor<string, []>("x2_31_end_0"), val = tensor<int32, [4]>([1, 16, 0, 64])];
            tensor<bool, [4]> x2_31_end_mask_0 = const()[name = tensor<string, []>("x2_31_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> x2_31_stride_0 = const()[name = tensor<string, []>("x2_31_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 16, ?, 32]> x2_31_cast_fp16 = slice_by_index(begin = x2_31_begin_0, end = x2_31_end_0, end_mask = x2_31_end_mask_0, stride = x2_31_stride_0, x = x_73_cast_fp16)[name = tensor<string, []>("x2_31_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_839_cast_fp16 = mul(x = x1_31_cast_fp16, y = blocks_0_att_head_cos_to_fp16)[name = tensor<string, []>("op_839_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_840_cast_fp16 = mul(x = x2_31_cast_fp16, y = blocks_0_att_head_sin_to_fp16)[name = tensor<string, []>("op_840_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_841_cast_fp16 = sub(x = var_839_cast_fp16, y = var_840_cast_fp16)[name = tensor<string, []>("op_841_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_842_cast_fp16 = mul(x = x2_31_cast_fp16, y = blocks_0_att_head_cos_to_fp16)[name = tensor<string, []>("op_842_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_843_cast_fp16 = mul(x = x1_31_cast_fp16, y = blocks_0_att_head_sin_to_fp16)[name = tensor<string, []>("op_843_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_844_cast_fp16 = add(x = var_842_cast_fp16, y = var_843_cast_fp16)[name = tensor<string, []>("op_844_cast_fp16")];
            tensor<int32, []> out_31_axis_0 = const()[name = tensor<string, []>("out_31_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 16, ?, 32, 2]> out_31_cast_fp16 = stack(axis = out_31_axis_0, values = (var_841_cast_fp16, var_844_cast_fp16))[name = tensor<string, []>("out_31_cast_fp16")];
            tensor<int32, [4]> concat_19x = const()[name = tensor<string, []>("concat_19x"), val = tensor<int32, [4]>([1, 16, -1, 64])];
            tensor<fp16, [1, 16, ?, 64]> k_31_cast_fp16 = reshape(shape = concat_19x, x = out_31_cast_fp16)[name = tensor<string, []>("k_31_cast_fp16")];
            tensor<fp16, []> mul_27_y_0_to_fp16 = const()[name = tensor<string, []>("mul_27_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 16, ?, 64]> mul_27_cast_fp16 = mul(x = q_31_cast_fp16, y = mul_27_y_0_to_fp16)[name = tensor<string, []>("mul_27_cast_fp16")];
            tensor<bool, []> matmul_7_transpose_y_0 = const()[name = tensor<string, []>("matmul_7_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_7_transpose_x_0 = const()[name = tensor<string, []>("matmul_7_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 16, ?, ?]> matmul_7_cast_fp16 = matmul(transpose_x = matmul_7_transpose_x_0, transpose_y = matmul_7_transpose_y_0, x = mul_27_cast_fp16, y = k_31_cast_fp16)[name = tensor<string, []>("matmul_7_cast_fp16")];
            tensor<string, []> matmul_7_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("matmul_7_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<int32, []> softmax_7_axis_0 = const()[name = tensor<string, []>("softmax_7_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp32, [1, 16, ?, ?]> matmul_7_cast_fp16_to_fp32 = cast(dtype = matmul_7_cast_fp16_to_fp32_dtype_0, x = matmul_7_cast_fp16)[name = tensor<string, []>("cast_144")];
            tensor<fp32, [1, 16, ?, ?]> softmax_7 = softmax(axis = softmax_7_axis_0, x = matmul_7_cast_fp16_to_fp32)[name = tensor<string, []>("softmax_7")];
            tensor<bool, []> y_15_transpose_x_0 = const()[name = tensor<string, []>("y_15_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> y_15_transpose_y_0 = const()[name = tensor<string, []>("y_15_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<string, []> softmax_7_to_fp16_dtype_0 = const()[name = tensor<string, []>("softmax_7_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 16, ?, ?]> softmax_7_to_fp16 = cast(dtype = softmax_7_to_fp16_dtype_0, x = softmax_7)[name = tensor<string, []>("cast_143")];
            tensor<fp16, [1, 16, ?, 64]> v_31_cast_fp16 = transpose(perm = v_31_perm_0, x = var_824_cast_fp16)[name = tensor<string, []>("transpose_20")];
            tensor<fp16, [1, 16, ?, 64]> y_15_cast_fp16 = matmul(transpose_x = y_15_transpose_x_0, transpose_y = y_15_transpose_y_0, x = softmax_7_to_fp16, y = v_31_cast_fp16)[name = tensor<string, []>("y_15_cast_fp16")];
            tensor<int32, [4]> var_849_perm_0 = const()[name = tensor<string, []>("op_849_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_850 = const()[name = tensor<string, []>("op_850"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, ?, 16, 64]> var_849_cast_fp16 = transpose(perm = var_849_perm_0, x = y_15_cast_fp16)[name = tensor<string, []>("transpose_19")];
            tensor<fp16, [1, ?, 1024]> input_101_cast_fp16 = reshape(shape = var_850, x = var_849_cast_fp16)[name = tensor<string, []>("input_101_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_7_att_c_proj_weight_to_fp16 = const()[name = tensor<string, []>("blocks_7_att_c_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(227668672)))];
            tensor<fp16, [1, ?, 1024]> linear_31_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = blocks_7_att_c_proj_weight_to_fp16, x = input_101_cast_fp16)[name = tensor<string, []>("linear_31_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_75_cast_fp16 = add(x = x_69_cast_fp16, y = linear_31_cast_fp16)[name = tensor<string, []>("x_75_cast_fp16")];
            tensor<string, []> x_75_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_75_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_789_promoted_1 = const()[name = tensor<string, []>("op_789_promoted_1"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, ?, 1024]> x_75_cast_fp16_to_fp32 = cast(dtype = x_75_cast_fp16_to_fp32_dtype_0, x = x_75_cast_fp16)[name = tensor<string, []>("cast_142")];
            tensor<fp32, [1, ?, 1024]> var_856 = pow(x = x_75_cast_fp16_to_fp32, y = var_789_promoted_1)[name = tensor<string, []>("op_856")];
            tensor<int32, [1]> norm_x_31_axes_0 = const()[name = tensor<string, []>("norm_x_31_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> norm_x_31_keep_dims_0 = const()[name = tensor<string, []>("norm_x_31_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, ?, 1]> norm_x_31 = reduce_mean(axes = norm_x_31_axes_0, keep_dims = norm_x_31_keep_dims_0, x = var_856)[name = tensor<string, []>("norm_x_31")];
            tensor<string, []> norm_x_31_to_fp16_dtype_0 = const()[name = tensor<string, []>("norm_x_31_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_859_to_fp16 = const()[name = tensor<string, []>("op_859_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> norm_x_31_to_fp16 = cast(dtype = norm_x_31_to_fp16_dtype_0, x = norm_x_31)[name = tensor<string, []>("cast_141")];
            tensor<fp16, [1, ?, 1]> var_860_cast_fp16 = add(x = norm_x_31_to_fp16, y = var_859_to_fp16)[name = tensor<string, []>("op_860_cast_fp16")];
            tensor<string, []> var_860_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_860_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_861_epsilon_0 = const()[name = tensor<string, []>("op_861_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, ?, 1]> var_860_cast_fp16_to_fp32 = cast(dtype = var_860_cast_fp16_to_fp32_dtype_0, x = var_860_cast_fp16)[name = tensor<string, []>("cast_140")];
            tensor<fp32, [1, ?, 1]> var_861 = rsqrt(epsilon = var_861_epsilon_0, x = var_860_cast_fp16_to_fp32)[name = tensor<string, []>("op_861")];
            tensor<string, []> var_861_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_861_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, ?, 1]> var_861_to_fp16 = cast(dtype = var_861_to_fp16_dtype_0, x = var_861)[name = tensor<string, []>("cast_139")];
            tensor<fp16, [1, ?, 1024]> var_862_cast_fp16 = mul(x = x_75_cast_fp16, y = var_861_to_fp16)[name = tensor<string, []>("op_862_cast_fp16")];
            tensor<fp16, [1024]> blocks_7_ffn_norm_weight_to_fp16 = const()[name = tensor<string, []>("blocks_7_ffn_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(229765888)))];
            tensor<fp16, [1, ?, 1024]> input_103_cast_fp16 = mul(x = var_862_cast_fp16, y = blocks_7_ffn_norm_weight_to_fp16)[name = tensor<string, []>("input_103_cast_fp16")];
            tensor<fp16, [4096, 1024]> blocks_7_mlp_fc1_weight_to_fp16 = const()[name = tensor<string, []>("blocks_7_mlp_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(229768000)))];
            tensor<fp16, [1, ?, 4096]> linear_32_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_7_mlp_fc1_weight_to_fp16, x = input_103_cast_fp16)[name = tensor<string, []>("linear_32_cast_fp16")];
            tensor<fp16, [1, ?, 4096]> input_107_cast_fp16 = silu(x = linear_32_cast_fp16)[name = tensor<string, []>("input_107_cast_fp16")];
            tensor<fp16, [1024, 4096]> blocks_7_mlp_fc2_weight_to_fp16 = const()[name = tensor<string, []>("blocks_7_mlp_fc2_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(238156672)))];
            tensor<fp16, [1, ?, 1024]> linear_33_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = blocks_7_mlp_fc2_weight_to_fp16, x = input_107_cast_fp16)[name = tensor<string, []>("linear_33_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_77_cast_fp16 = add(x = x_75_cast_fp16, y = linear_33_cast_fp16)[name = tensor<string, []>("x_77_cast_fp16")];
            tensor<string, []> x_77_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_77_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_881_promoted = const()[name = tensor<string, []>("op_881_promoted"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, ?, 1024]> x_77_cast_fp16_to_fp32 = cast(dtype = x_77_cast_fp16_to_fp32_dtype_0, x = x_77_cast_fp16)[name = tensor<string, []>("cast_138")];
            tensor<fp32, [1, ?, 1024]> var_891 = pow(x = x_77_cast_fp16_to_fp32, y = var_881_promoted)[name = tensor<string, []>("op_891")];
            tensor<int32, [1]> norm_x_33_axes_0 = const()[name = tensor<string, []>("norm_x_33_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> norm_x_33_keep_dims_0 = const()[name = tensor<string, []>("norm_x_33_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, ?, 1]> norm_x_33 = reduce_mean(axes = norm_x_33_axes_0, keep_dims = norm_x_33_keep_dims_0, x = var_891)[name = tensor<string, []>("norm_x_33")];
            tensor<string, []> norm_x_33_to_fp16_dtype_0 = const()[name = tensor<string, []>("norm_x_33_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_894_to_fp16 = const()[name = tensor<string, []>("op_894_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> norm_x_33_to_fp16 = cast(dtype = norm_x_33_to_fp16_dtype_0, x = norm_x_33)[name = tensor<string, []>("cast_137")];
            tensor<fp16, [1, ?, 1]> var_895_cast_fp16 = add(x = norm_x_33_to_fp16, y = var_894_to_fp16)[name = tensor<string, []>("op_895_cast_fp16")];
            tensor<string, []> var_895_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_895_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_896_epsilon_0 = const()[name = tensor<string, []>("op_896_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, ?, 1]> var_895_cast_fp16_to_fp32 = cast(dtype = var_895_cast_fp16_to_fp32_dtype_0, x = var_895_cast_fp16)[name = tensor<string, []>("cast_136")];
            tensor<fp32, [1, ?, 1]> var_896 = rsqrt(epsilon = var_896_epsilon_0, x = var_895_cast_fp16_to_fp32)[name = tensor<string, []>("op_896")];
            tensor<string, []> var_896_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_896_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, ?, 1]> var_896_to_fp16 = cast(dtype = var_896_to_fp16_dtype_0, x = var_896)[name = tensor<string, []>("cast_135")];
            tensor<fp16, [1, ?, 1024]> var_897_cast_fp16 = mul(x = x_77_cast_fp16, y = var_896_to_fp16)[name = tensor<string, []>("op_897_cast_fp16")];
            tensor<fp16, [1024]> blocks_8_att_norm_weight_to_fp16 = const()[name = tensor<string, []>("blocks_8_att_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(246545344)))];
            tensor<fp16, [1, ?, 1024]> input_109_cast_fp16 = mul(x = var_897_cast_fp16, y = blocks_8_att_norm_weight_to_fp16)[name = tensor<string, []>("input_109_cast_fp16")];
            tensor<fp16, [3072, 1024]> blocks_8_att_c_attn_weight_to_fp16 = const()[name = tensor<string, []>("blocks_8_att_c_attn_weight_to_fp16"), val = tensor<fp16, [3072, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(246547456)))];
            tensor<fp16, [1, ?, 3072]> linear_34_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = blocks_8_att_c_attn_weight_to_fp16, x = input_109_cast_fp16)[name = tensor<string, []>("linear_34_cast_fp16")];
            tensor<int32, [3]> tile_8 = const()[name = tensor<string, []>("tile_8"), val = tensor<int32, [3]>([1024, 1024, 1024])];
            tensor<int32, []> var_905_axis_0 = const()[name = tensor<string, []>("op_905_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, ?, 1024]> var_905_cast_fp16_0, tensor<fp16, [1, ?, 1024]> var_905_cast_fp16_1, tensor<fp16, [1, ?, 1024]> var_905_cast_fp16_2 = split(axis = var_905_axis_0, split_sizes = tile_8, x = linear_34_cast_fp16)[name = tensor<string, []>("op_905_cast_fp16")];
            tensor<int32, [4]> var_909 = const()[name = tensor<string, []>("op_909"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_910_cast_fp16 = reshape(shape = var_909, x = var_905_cast_fp16_0)[name = tensor<string, []>("op_910_cast_fp16")];
            tensor<int32, [4]> x_79_perm_0 = const()[name = tensor<string, []>("x_79_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_912 = const()[name = tensor<string, []>("op_912"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_913_cast_fp16 = reshape(shape = var_912, x = var_905_cast_fp16_1)[name = tensor<string, []>("op_913_cast_fp16")];
            tensor<int32, [4]> x_81_perm_0 = const()[name = tensor<string, []>("x_81_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_915 = const()[name = tensor<string, []>("op_915"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_916_cast_fp16 = reshape(shape = var_915, x = var_905_cast_fp16_2)[name = tensor<string, []>("op_916_cast_fp16")];
            tensor<int32, [4]> v_35_perm_0 = const()[name = tensor<string, []>("v_35_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> x1_33_begin_0 = const()[name = tensor<string, []>("x1_33_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_33_end_0 = const()[name = tensor<string, []>("x1_33_end_0"), val = tensor<int32, [4]>([1, 16, 0, 64])];
            tensor<bool, [4]> x1_33_end_mask_0 = const()[name = tensor<string, []>("x1_33_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> x1_33_stride_0 = const()[name = tensor<string, []>("x1_33_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 16, ?, 64]> x_79_cast_fp16 = transpose(perm = x_79_perm_0, x = var_910_cast_fp16)[name = tensor<string, []>("transpose_18")];
            tensor<fp16, [1, 16, ?, 32]> x1_33_cast_fp16 = slice_by_index(begin = x1_33_begin_0, end = x1_33_end_0, end_mask = x1_33_end_mask_0, stride = x1_33_stride_0, x = x_79_cast_fp16)[name = tensor<string, []>("x1_33_cast_fp16")];
            tensor<int32, [4]> x2_33_begin_0 = const()[name = tensor<string, []>("x2_33_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 1])];
            tensor<int32, [4]> x2_33_end_0 = const()[name = tensor<string, []>("x2_33_end_0"), val = tensor<int32, [4]>([1, 16, 0, 64])];
            tensor<bool, [4]> x2_33_end_mask_0 = const()[name = tensor<string, []>("x2_33_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> x2_33_stride_0 = const()[name = tensor<string, []>("x2_33_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 16, ?, 32]> x2_33_cast_fp16 = slice_by_index(begin = x2_33_begin_0, end = x2_33_end_0, end_mask = x2_33_end_mask_0, stride = x2_33_stride_0, x = x_79_cast_fp16)[name = tensor<string, []>("x2_33_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_920_cast_fp16 = mul(x = x1_33_cast_fp16, y = blocks_0_att_head_cos_to_fp16)[name = tensor<string, []>("op_920_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_921_cast_fp16 = mul(x = x2_33_cast_fp16, y = blocks_0_att_head_sin_to_fp16)[name = tensor<string, []>("op_921_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_922_cast_fp16 = sub(x = var_920_cast_fp16, y = var_921_cast_fp16)[name = tensor<string, []>("op_922_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_923_cast_fp16 = mul(x = x2_33_cast_fp16, y = blocks_0_att_head_cos_to_fp16)[name = tensor<string, []>("op_923_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_924_cast_fp16 = mul(x = x1_33_cast_fp16, y = blocks_0_att_head_sin_to_fp16)[name = tensor<string, []>("op_924_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_925_cast_fp16 = add(x = var_923_cast_fp16, y = var_924_cast_fp16)[name = tensor<string, []>("op_925_cast_fp16")];
            tensor<int32, []> out_33_axis_0 = const()[name = tensor<string, []>("out_33_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 16, ?, 32, 2]> out_33_cast_fp16 = stack(axis = out_33_axis_0, values = (var_922_cast_fp16, var_925_cast_fp16))[name = tensor<string, []>("out_33_cast_fp16")];
            tensor<int32, [4]> concat_20x = const()[name = tensor<string, []>("concat_20x"), val = tensor<int32, [4]>([1, 16, -1, 64])];
            tensor<fp16, [1, 16, ?, 64]> q_35_cast_fp16 = reshape(shape = concat_20x, x = out_33_cast_fp16)[name = tensor<string, []>("q_35_cast_fp16")];
            tensor<int32, [4]> x1_35_begin_0 = const()[name = tensor<string, []>("x1_35_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_35_end_0 = const()[name = tensor<string, []>("x1_35_end_0"), val = tensor<int32, [4]>([1, 16, 0, 64])];
            tensor<bool, [4]> x1_35_end_mask_0 = const()[name = tensor<string, []>("x1_35_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> x1_35_stride_0 = const()[name = tensor<string, []>("x1_35_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 16, ?, 64]> x_81_cast_fp16 = transpose(perm = x_81_perm_0, x = var_913_cast_fp16)[name = tensor<string, []>("transpose_17")];
            tensor<fp16, [1, 16, ?, 32]> x1_35_cast_fp16 = slice_by_index(begin = x1_35_begin_0, end = x1_35_end_0, end_mask = x1_35_end_mask_0, stride = x1_35_stride_0, x = x_81_cast_fp16)[name = tensor<string, []>("x1_35_cast_fp16")];
            tensor<int32, [4]> x2_35_begin_0 = const()[name = tensor<string, []>("x2_35_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 1])];
            tensor<int32, [4]> x2_35_end_0 = const()[name = tensor<string, []>("x2_35_end_0"), val = tensor<int32, [4]>([1, 16, 0, 64])];
            tensor<bool, [4]> x2_35_end_mask_0 = const()[name = tensor<string, []>("x2_35_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> x2_35_stride_0 = const()[name = tensor<string, []>("x2_35_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 16, ?, 32]> x2_35_cast_fp16 = slice_by_index(begin = x2_35_begin_0, end = x2_35_end_0, end_mask = x2_35_end_mask_0, stride = x2_35_stride_0, x = x_81_cast_fp16)[name = tensor<string, []>("x2_35_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_931_cast_fp16 = mul(x = x1_35_cast_fp16, y = blocks_0_att_head_cos_to_fp16)[name = tensor<string, []>("op_931_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_932_cast_fp16 = mul(x = x2_35_cast_fp16, y = blocks_0_att_head_sin_to_fp16)[name = tensor<string, []>("op_932_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_933_cast_fp16 = sub(x = var_931_cast_fp16, y = var_932_cast_fp16)[name = tensor<string, []>("op_933_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_934_cast_fp16 = mul(x = x2_35_cast_fp16, y = blocks_0_att_head_cos_to_fp16)[name = tensor<string, []>("op_934_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_935_cast_fp16 = mul(x = x1_35_cast_fp16, y = blocks_0_att_head_sin_to_fp16)[name = tensor<string, []>("op_935_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_936_cast_fp16 = add(x = var_934_cast_fp16, y = var_935_cast_fp16)[name = tensor<string, []>("op_936_cast_fp16")];
            tensor<int32, []> out_35_axis_0 = const()[name = tensor<string, []>("out_35_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 16, ?, 32, 2]> out_35_cast_fp16 = stack(axis = out_35_axis_0, values = (var_933_cast_fp16, var_936_cast_fp16))[name = tensor<string, []>("out_35_cast_fp16")];
            tensor<int32, [4]> concat_21x = const()[name = tensor<string, []>("concat_21x"), val = tensor<int32, [4]>([1, 16, -1, 64])];
            tensor<fp16, [1, 16, ?, 64]> k_35_cast_fp16 = reshape(shape = concat_21x, x = out_35_cast_fp16)[name = tensor<string, []>("k_35_cast_fp16")];
            tensor<fp16, []> mul_30_y_0_to_fp16 = const()[name = tensor<string, []>("mul_30_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 16, ?, 64]> mul_30_cast_fp16 = mul(x = q_35_cast_fp16, y = mul_30_y_0_to_fp16)[name = tensor<string, []>("mul_30_cast_fp16")];
            tensor<bool, []> matmul_8_transpose_y_0 = const()[name = tensor<string, []>("matmul_8_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_8_transpose_x_0 = const()[name = tensor<string, []>("matmul_8_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 16, ?, ?]> matmul_8_cast_fp16 = matmul(transpose_x = matmul_8_transpose_x_0, transpose_y = matmul_8_transpose_y_0, x = mul_30_cast_fp16, y = k_35_cast_fp16)[name = tensor<string, []>("matmul_8_cast_fp16")];
            tensor<string, []> matmul_8_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("matmul_8_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<int32, []> softmax_8_axis_0 = const()[name = tensor<string, []>("softmax_8_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp32, [1, 16, ?, ?]> matmul_8_cast_fp16_to_fp32 = cast(dtype = matmul_8_cast_fp16_to_fp32_dtype_0, x = matmul_8_cast_fp16)[name = tensor<string, []>("cast_134")];
            tensor<fp32, [1, 16, ?, ?]> softmax_8 = softmax(axis = softmax_8_axis_0, x = matmul_8_cast_fp16_to_fp32)[name = tensor<string, []>("softmax_8")];
            tensor<bool, []> y_17_transpose_x_0 = const()[name = tensor<string, []>("y_17_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> y_17_transpose_y_0 = const()[name = tensor<string, []>("y_17_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<string, []> softmax_8_to_fp16_dtype_0 = const()[name = tensor<string, []>("softmax_8_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 16, ?, ?]> softmax_8_to_fp16 = cast(dtype = softmax_8_to_fp16_dtype_0, x = softmax_8)[name = tensor<string, []>("cast_133")];
            tensor<fp16, [1, 16, ?, 64]> v_35_cast_fp16 = transpose(perm = v_35_perm_0, x = var_916_cast_fp16)[name = tensor<string, []>("transpose_16")];
            tensor<fp16, [1, 16, ?, 64]> y_17_cast_fp16 = matmul(transpose_x = y_17_transpose_x_0, transpose_y = y_17_transpose_y_0, x = softmax_8_to_fp16, y = v_35_cast_fp16)[name = tensor<string, []>("y_17_cast_fp16")];
            tensor<int32, [4]> var_941_perm_0 = const()[name = tensor<string, []>("op_941_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_942 = const()[name = tensor<string, []>("op_942"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, ?, 16, 64]> var_941_cast_fp16 = transpose(perm = var_941_perm_0, x = y_17_cast_fp16)[name = tensor<string, []>("transpose_15")];
            tensor<fp16, [1, ?, 1024]> input_111_cast_fp16 = reshape(shape = var_942, x = var_941_cast_fp16)[name = tensor<string, []>("input_111_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_8_att_c_proj_weight_to_fp16 = const()[name = tensor<string, []>("blocks_8_att_c_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(252838976)))];
            tensor<fp16, [1, ?, 1024]> linear_35_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = blocks_8_att_c_proj_weight_to_fp16, x = input_111_cast_fp16)[name = tensor<string, []>("linear_35_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_83_cast_fp16 = add(x = x_77_cast_fp16, y = linear_35_cast_fp16)[name = tensor<string, []>("x_83_cast_fp16")];
            tensor<string, []> x_83_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_83_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_881_promoted_1 = const()[name = tensor<string, []>("op_881_promoted_1"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, ?, 1024]> x_83_cast_fp16_to_fp32 = cast(dtype = x_83_cast_fp16_to_fp32_dtype_0, x = x_83_cast_fp16)[name = tensor<string, []>("cast_132")];
            tensor<fp32, [1, ?, 1024]> var_948 = pow(x = x_83_cast_fp16_to_fp32, y = var_881_promoted_1)[name = tensor<string, []>("op_948")];
            tensor<int32, [1]> norm_x_35_axes_0 = const()[name = tensor<string, []>("norm_x_35_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> norm_x_35_keep_dims_0 = const()[name = tensor<string, []>("norm_x_35_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, ?, 1]> norm_x_35 = reduce_mean(axes = norm_x_35_axes_0, keep_dims = norm_x_35_keep_dims_0, x = var_948)[name = tensor<string, []>("norm_x_35")];
            tensor<string, []> norm_x_35_to_fp16_dtype_0 = const()[name = tensor<string, []>("norm_x_35_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_951_to_fp16 = const()[name = tensor<string, []>("op_951_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> norm_x_35_to_fp16 = cast(dtype = norm_x_35_to_fp16_dtype_0, x = norm_x_35)[name = tensor<string, []>("cast_131")];
            tensor<fp16, [1, ?, 1]> var_952_cast_fp16 = add(x = norm_x_35_to_fp16, y = var_951_to_fp16)[name = tensor<string, []>("op_952_cast_fp16")];
            tensor<string, []> var_952_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_952_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_953_epsilon_0 = const()[name = tensor<string, []>("op_953_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, ?, 1]> var_952_cast_fp16_to_fp32 = cast(dtype = var_952_cast_fp16_to_fp32_dtype_0, x = var_952_cast_fp16)[name = tensor<string, []>("cast_130")];
            tensor<fp32, [1, ?, 1]> var_953 = rsqrt(epsilon = var_953_epsilon_0, x = var_952_cast_fp16_to_fp32)[name = tensor<string, []>("op_953")];
            tensor<string, []> var_953_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_953_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, ?, 1]> var_953_to_fp16 = cast(dtype = var_953_to_fp16_dtype_0, x = var_953)[name = tensor<string, []>("cast_129")];
            tensor<fp16, [1, ?, 1024]> var_954_cast_fp16 = mul(x = x_83_cast_fp16, y = var_953_to_fp16)[name = tensor<string, []>("op_954_cast_fp16")];
            tensor<fp16, [1024]> blocks_8_ffn_norm_weight_to_fp16 = const()[name = tensor<string, []>("blocks_8_ffn_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(254936192)))];
            tensor<fp16, [1, ?, 1024]> input_113_cast_fp16 = mul(x = var_954_cast_fp16, y = blocks_8_ffn_norm_weight_to_fp16)[name = tensor<string, []>("input_113_cast_fp16")];
            tensor<fp16, [4096, 1024]> blocks_8_mlp_fc1_weight_to_fp16 = const()[name = tensor<string, []>("blocks_8_mlp_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(254938304)))];
            tensor<fp16, [1, ?, 4096]> linear_36_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_8_mlp_fc1_weight_to_fp16, x = input_113_cast_fp16)[name = tensor<string, []>("linear_36_cast_fp16")];
            tensor<fp16, [1, ?, 4096]> input_117_cast_fp16 = silu(x = linear_36_cast_fp16)[name = tensor<string, []>("input_117_cast_fp16")];
            tensor<fp16, [1024, 4096]> blocks_8_mlp_fc2_weight_to_fp16 = const()[name = tensor<string, []>("blocks_8_mlp_fc2_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(263326976)))];
            tensor<fp16, [1, ?, 1024]> linear_37_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = blocks_8_mlp_fc2_weight_to_fp16, x = input_117_cast_fp16)[name = tensor<string, []>("linear_37_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_85_cast_fp16 = add(x = x_83_cast_fp16, y = linear_37_cast_fp16)[name = tensor<string, []>("x_85_cast_fp16")];
            tensor<string, []> x_85_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_85_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_973_promoted = const()[name = tensor<string, []>("op_973_promoted"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, ?, 1024]> x_85_cast_fp16_to_fp32 = cast(dtype = x_85_cast_fp16_to_fp32_dtype_0, x = x_85_cast_fp16)[name = tensor<string, []>("cast_128")];
            tensor<fp32, [1, ?, 1024]> var_983 = pow(x = x_85_cast_fp16_to_fp32, y = var_973_promoted)[name = tensor<string, []>("op_983")];
            tensor<int32, [1]> norm_x_37_axes_0 = const()[name = tensor<string, []>("norm_x_37_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> norm_x_37_keep_dims_0 = const()[name = tensor<string, []>("norm_x_37_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, ?, 1]> norm_x_37 = reduce_mean(axes = norm_x_37_axes_0, keep_dims = norm_x_37_keep_dims_0, x = var_983)[name = tensor<string, []>("norm_x_37")];
            tensor<string, []> norm_x_37_to_fp16_dtype_0 = const()[name = tensor<string, []>("norm_x_37_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_986_to_fp16 = const()[name = tensor<string, []>("op_986_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> norm_x_37_to_fp16 = cast(dtype = norm_x_37_to_fp16_dtype_0, x = norm_x_37)[name = tensor<string, []>("cast_127")];
            tensor<fp16, [1, ?, 1]> var_987_cast_fp16 = add(x = norm_x_37_to_fp16, y = var_986_to_fp16)[name = tensor<string, []>("op_987_cast_fp16")];
            tensor<string, []> var_987_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_987_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_988_epsilon_0 = const()[name = tensor<string, []>("op_988_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, ?, 1]> var_987_cast_fp16_to_fp32 = cast(dtype = var_987_cast_fp16_to_fp32_dtype_0, x = var_987_cast_fp16)[name = tensor<string, []>("cast_126")];
            tensor<fp32, [1, ?, 1]> var_988 = rsqrt(epsilon = var_988_epsilon_0, x = var_987_cast_fp16_to_fp32)[name = tensor<string, []>("op_988")];
            tensor<string, []> var_988_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_988_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, ?, 1]> var_988_to_fp16 = cast(dtype = var_988_to_fp16_dtype_0, x = var_988)[name = tensor<string, []>("cast_125")];
            tensor<fp16, [1, ?, 1024]> var_989_cast_fp16 = mul(x = x_85_cast_fp16, y = var_988_to_fp16)[name = tensor<string, []>("op_989_cast_fp16")];
            tensor<fp16, [1024]> blocks_9_att_norm_weight_to_fp16 = const()[name = tensor<string, []>("blocks_9_att_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(271715648)))];
            tensor<fp16, [1, ?, 1024]> input_119_cast_fp16 = mul(x = var_989_cast_fp16, y = blocks_9_att_norm_weight_to_fp16)[name = tensor<string, []>("input_119_cast_fp16")];
            tensor<fp16, [3072, 1024]> blocks_9_att_c_attn_weight_to_fp16 = const()[name = tensor<string, []>("blocks_9_att_c_attn_weight_to_fp16"), val = tensor<fp16, [3072, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(271717760)))];
            tensor<fp16, [1, ?, 3072]> linear_38_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = blocks_9_att_c_attn_weight_to_fp16, x = input_119_cast_fp16)[name = tensor<string, []>("linear_38_cast_fp16")];
            tensor<int32, [3]> tile_9 = const()[name = tensor<string, []>("tile_9"), val = tensor<int32, [3]>([1024, 1024, 1024])];
            tensor<int32, []> var_997_axis_0 = const()[name = tensor<string, []>("op_997_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, ?, 1024]> var_997_cast_fp16_0, tensor<fp16, [1, ?, 1024]> var_997_cast_fp16_1, tensor<fp16, [1, ?, 1024]> var_997_cast_fp16_2 = split(axis = var_997_axis_0, split_sizes = tile_9, x = linear_38_cast_fp16)[name = tensor<string, []>("op_997_cast_fp16")];
            tensor<int32, [4]> var_1001 = const()[name = tensor<string, []>("op_1001"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_1002_cast_fp16 = reshape(shape = var_1001, x = var_997_cast_fp16_0)[name = tensor<string, []>("op_1002_cast_fp16")];
            tensor<int32, [4]> x_87_perm_0 = const()[name = tensor<string, []>("x_87_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_1004 = const()[name = tensor<string, []>("op_1004"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_1005_cast_fp16 = reshape(shape = var_1004, x = var_997_cast_fp16_1)[name = tensor<string, []>("op_1005_cast_fp16")];
            tensor<int32, [4]> x_89_perm_0 = const()[name = tensor<string, []>("x_89_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_1007 = const()[name = tensor<string, []>("op_1007"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_1008_cast_fp16 = reshape(shape = var_1007, x = var_997_cast_fp16_2)[name = tensor<string, []>("op_1008_cast_fp16")];
            tensor<int32, [4]> v_39_perm_0 = const()[name = tensor<string, []>("v_39_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> x1_37_begin_0 = const()[name = tensor<string, []>("x1_37_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_37_end_0 = const()[name = tensor<string, []>("x1_37_end_0"), val = tensor<int32, [4]>([1, 16, 0, 64])];
            tensor<bool, [4]> x1_37_end_mask_0 = const()[name = tensor<string, []>("x1_37_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> x1_37_stride_0 = const()[name = tensor<string, []>("x1_37_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 16, ?, 64]> x_87_cast_fp16 = transpose(perm = x_87_perm_0, x = var_1002_cast_fp16)[name = tensor<string, []>("transpose_14")];
            tensor<fp16, [1, 16, ?, 32]> x1_37_cast_fp16 = slice_by_index(begin = x1_37_begin_0, end = x1_37_end_0, end_mask = x1_37_end_mask_0, stride = x1_37_stride_0, x = x_87_cast_fp16)[name = tensor<string, []>("x1_37_cast_fp16")];
            tensor<int32, [4]> x2_37_begin_0 = const()[name = tensor<string, []>("x2_37_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 1])];
            tensor<int32, [4]> x2_37_end_0 = const()[name = tensor<string, []>("x2_37_end_0"), val = tensor<int32, [4]>([1, 16, 0, 64])];
            tensor<bool, [4]> x2_37_end_mask_0 = const()[name = tensor<string, []>("x2_37_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> x2_37_stride_0 = const()[name = tensor<string, []>("x2_37_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 16, ?, 32]> x2_37_cast_fp16 = slice_by_index(begin = x2_37_begin_0, end = x2_37_end_0, end_mask = x2_37_end_mask_0, stride = x2_37_stride_0, x = x_87_cast_fp16)[name = tensor<string, []>("x2_37_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_1012_cast_fp16 = mul(x = x1_37_cast_fp16, y = blocks_0_att_head_cos_to_fp16)[name = tensor<string, []>("op_1012_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_1013_cast_fp16 = mul(x = x2_37_cast_fp16, y = blocks_0_att_head_sin_to_fp16)[name = tensor<string, []>("op_1013_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_1014_cast_fp16 = sub(x = var_1012_cast_fp16, y = var_1013_cast_fp16)[name = tensor<string, []>("op_1014_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_1015_cast_fp16 = mul(x = x2_37_cast_fp16, y = blocks_0_att_head_cos_to_fp16)[name = tensor<string, []>("op_1015_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_1016_cast_fp16 = mul(x = x1_37_cast_fp16, y = blocks_0_att_head_sin_to_fp16)[name = tensor<string, []>("op_1016_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_1017_cast_fp16 = add(x = var_1015_cast_fp16, y = var_1016_cast_fp16)[name = tensor<string, []>("op_1017_cast_fp16")];
            tensor<int32, []> out_37_axis_0 = const()[name = tensor<string, []>("out_37_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 16, ?, 32, 2]> out_37_cast_fp16 = stack(axis = out_37_axis_0, values = (var_1014_cast_fp16, var_1017_cast_fp16))[name = tensor<string, []>("out_37_cast_fp16")];
            tensor<int32, [4]> concat_22x = const()[name = tensor<string, []>("concat_22x"), val = tensor<int32, [4]>([1, 16, -1, 64])];
            tensor<fp16, [1, 16, ?, 64]> q_39_cast_fp16 = reshape(shape = concat_22x, x = out_37_cast_fp16)[name = tensor<string, []>("q_39_cast_fp16")];
            tensor<int32, [4]> x1_39_begin_0 = const()[name = tensor<string, []>("x1_39_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_39_end_0 = const()[name = tensor<string, []>("x1_39_end_0"), val = tensor<int32, [4]>([1, 16, 0, 64])];
            tensor<bool, [4]> x1_39_end_mask_0 = const()[name = tensor<string, []>("x1_39_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> x1_39_stride_0 = const()[name = tensor<string, []>("x1_39_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 16, ?, 64]> x_89_cast_fp16 = transpose(perm = x_89_perm_0, x = var_1005_cast_fp16)[name = tensor<string, []>("transpose_13")];
            tensor<fp16, [1, 16, ?, 32]> x1_39_cast_fp16 = slice_by_index(begin = x1_39_begin_0, end = x1_39_end_0, end_mask = x1_39_end_mask_0, stride = x1_39_stride_0, x = x_89_cast_fp16)[name = tensor<string, []>("x1_39_cast_fp16")];
            tensor<int32, [4]> x2_39_begin_0 = const()[name = tensor<string, []>("x2_39_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 1])];
            tensor<int32, [4]> x2_39_end_0 = const()[name = tensor<string, []>("x2_39_end_0"), val = tensor<int32, [4]>([1, 16, 0, 64])];
            tensor<bool, [4]> x2_39_end_mask_0 = const()[name = tensor<string, []>("x2_39_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> x2_39_stride_0 = const()[name = tensor<string, []>("x2_39_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 16, ?, 32]> x2_39_cast_fp16 = slice_by_index(begin = x2_39_begin_0, end = x2_39_end_0, end_mask = x2_39_end_mask_0, stride = x2_39_stride_0, x = x_89_cast_fp16)[name = tensor<string, []>("x2_39_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_1023_cast_fp16 = mul(x = x1_39_cast_fp16, y = blocks_0_att_head_cos_to_fp16)[name = tensor<string, []>("op_1023_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_1024_cast_fp16 = mul(x = x2_39_cast_fp16, y = blocks_0_att_head_sin_to_fp16)[name = tensor<string, []>("op_1024_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_1025_cast_fp16 = sub(x = var_1023_cast_fp16, y = var_1024_cast_fp16)[name = tensor<string, []>("op_1025_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_1026_cast_fp16 = mul(x = x2_39_cast_fp16, y = blocks_0_att_head_cos_to_fp16)[name = tensor<string, []>("op_1026_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_1027_cast_fp16 = mul(x = x1_39_cast_fp16, y = blocks_0_att_head_sin_to_fp16)[name = tensor<string, []>("op_1027_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_1028_cast_fp16 = add(x = var_1026_cast_fp16, y = var_1027_cast_fp16)[name = tensor<string, []>("op_1028_cast_fp16")];
            tensor<int32, []> out_39_axis_0 = const()[name = tensor<string, []>("out_39_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 16, ?, 32, 2]> out_39_cast_fp16 = stack(axis = out_39_axis_0, values = (var_1025_cast_fp16, var_1028_cast_fp16))[name = tensor<string, []>("out_39_cast_fp16")];
            tensor<int32, [4]> concat_23x = const()[name = tensor<string, []>("concat_23x"), val = tensor<int32, [4]>([1, 16, -1, 64])];
            tensor<fp16, [1, 16, ?, 64]> k_39_cast_fp16 = reshape(shape = concat_23x, x = out_39_cast_fp16)[name = tensor<string, []>("k_39_cast_fp16")];
            tensor<fp16, []> mul_33_y_0_to_fp16 = const()[name = tensor<string, []>("mul_33_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 16, ?, 64]> mul_33_cast_fp16 = mul(x = q_39_cast_fp16, y = mul_33_y_0_to_fp16)[name = tensor<string, []>("mul_33_cast_fp16")];
            tensor<bool, []> matmul_9_transpose_y_0 = const()[name = tensor<string, []>("matmul_9_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_9_transpose_x_0 = const()[name = tensor<string, []>("matmul_9_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 16, ?, ?]> matmul_9_cast_fp16 = matmul(transpose_x = matmul_9_transpose_x_0, transpose_y = matmul_9_transpose_y_0, x = mul_33_cast_fp16, y = k_39_cast_fp16)[name = tensor<string, []>("matmul_9_cast_fp16")];
            tensor<string, []> matmul_9_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("matmul_9_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<int32, []> softmax_9_axis_0 = const()[name = tensor<string, []>("softmax_9_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp32, [1, 16, ?, ?]> matmul_9_cast_fp16_to_fp32 = cast(dtype = matmul_9_cast_fp16_to_fp32_dtype_0, x = matmul_9_cast_fp16)[name = tensor<string, []>("cast_124")];
            tensor<fp32, [1, 16, ?, ?]> softmax_9 = softmax(axis = softmax_9_axis_0, x = matmul_9_cast_fp16_to_fp32)[name = tensor<string, []>("softmax_9")];
            tensor<bool, []> y_19_transpose_x_0 = const()[name = tensor<string, []>("y_19_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> y_19_transpose_y_0 = const()[name = tensor<string, []>("y_19_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<string, []> softmax_9_to_fp16_dtype_0 = const()[name = tensor<string, []>("softmax_9_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 16, ?, ?]> softmax_9_to_fp16 = cast(dtype = softmax_9_to_fp16_dtype_0, x = softmax_9)[name = tensor<string, []>("cast_123")];
            tensor<fp16, [1, 16, ?, 64]> v_39_cast_fp16 = transpose(perm = v_39_perm_0, x = var_1008_cast_fp16)[name = tensor<string, []>("transpose_12")];
            tensor<fp16, [1, 16, ?, 64]> y_19_cast_fp16 = matmul(transpose_x = y_19_transpose_x_0, transpose_y = y_19_transpose_y_0, x = softmax_9_to_fp16, y = v_39_cast_fp16)[name = tensor<string, []>("y_19_cast_fp16")];
            tensor<int32, [4]> var_1033_perm_0 = const()[name = tensor<string, []>("op_1033_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1034 = const()[name = tensor<string, []>("op_1034"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, ?, 16, 64]> var_1033_cast_fp16 = transpose(perm = var_1033_perm_0, x = y_19_cast_fp16)[name = tensor<string, []>("transpose_11")];
            tensor<fp16, [1, ?, 1024]> input_121_cast_fp16 = reshape(shape = var_1034, x = var_1033_cast_fp16)[name = tensor<string, []>("input_121_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_9_att_c_proj_weight_to_fp16 = const()[name = tensor<string, []>("blocks_9_att_c_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(278009280)))];
            tensor<fp16, [1, ?, 1024]> linear_39_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = blocks_9_att_c_proj_weight_to_fp16, x = input_121_cast_fp16)[name = tensor<string, []>("linear_39_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_91_cast_fp16 = add(x = x_85_cast_fp16, y = linear_39_cast_fp16)[name = tensor<string, []>("x_91_cast_fp16")];
            tensor<string, []> x_91_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_91_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_973_promoted_1 = const()[name = tensor<string, []>("op_973_promoted_1"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, ?, 1024]> x_91_cast_fp16_to_fp32 = cast(dtype = x_91_cast_fp16_to_fp32_dtype_0, x = x_91_cast_fp16)[name = tensor<string, []>("cast_122")];
            tensor<fp32, [1, ?, 1024]> var_1040 = pow(x = x_91_cast_fp16_to_fp32, y = var_973_promoted_1)[name = tensor<string, []>("op_1040")];
            tensor<int32, [1]> norm_x_39_axes_0 = const()[name = tensor<string, []>("norm_x_39_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> norm_x_39_keep_dims_0 = const()[name = tensor<string, []>("norm_x_39_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, ?, 1]> norm_x_39 = reduce_mean(axes = norm_x_39_axes_0, keep_dims = norm_x_39_keep_dims_0, x = var_1040)[name = tensor<string, []>("norm_x_39")];
            tensor<string, []> norm_x_39_to_fp16_dtype_0 = const()[name = tensor<string, []>("norm_x_39_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_1043_to_fp16 = const()[name = tensor<string, []>("op_1043_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> norm_x_39_to_fp16 = cast(dtype = norm_x_39_to_fp16_dtype_0, x = norm_x_39)[name = tensor<string, []>("cast_121")];
            tensor<fp16, [1, ?, 1]> var_1044_cast_fp16 = add(x = norm_x_39_to_fp16, y = var_1043_to_fp16)[name = tensor<string, []>("op_1044_cast_fp16")];
            tensor<string, []> var_1044_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_1044_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_1045_epsilon_0 = const()[name = tensor<string, []>("op_1045_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, ?, 1]> var_1044_cast_fp16_to_fp32 = cast(dtype = var_1044_cast_fp16_to_fp32_dtype_0, x = var_1044_cast_fp16)[name = tensor<string, []>("cast_120")];
            tensor<fp32, [1, ?, 1]> var_1045 = rsqrt(epsilon = var_1045_epsilon_0, x = var_1044_cast_fp16_to_fp32)[name = tensor<string, []>("op_1045")];
            tensor<string, []> var_1045_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_1045_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, ?, 1]> var_1045_to_fp16 = cast(dtype = var_1045_to_fp16_dtype_0, x = var_1045)[name = tensor<string, []>("cast_119")];
            tensor<fp16, [1, ?, 1024]> var_1046_cast_fp16 = mul(x = x_91_cast_fp16, y = var_1045_to_fp16)[name = tensor<string, []>("op_1046_cast_fp16")];
            tensor<fp16, [1024]> blocks_9_ffn_norm_weight_to_fp16 = const()[name = tensor<string, []>("blocks_9_ffn_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(280106496)))];
            tensor<fp16, [1, ?, 1024]> input_123_cast_fp16 = mul(x = var_1046_cast_fp16, y = blocks_9_ffn_norm_weight_to_fp16)[name = tensor<string, []>("input_123_cast_fp16")];
            tensor<fp16, [4096, 1024]> blocks_9_mlp_fc1_weight_to_fp16 = const()[name = tensor<string, []>("blocks_9_mlp_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(280108608)))];
            tensor<fp16, [1, ?, 4096]> linear_40_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_9_mlp_fc1_weight_to_fp16, x = input_123_cast_fp16)[name = tensor<string, []>("linear_40_cast_fp16")];
            tensor<fp16, [1, ?, 4096]> input_127_cast_fp16 = silu(x = linear_40_cast_fp16)[name = tensor<string, []>("input_127_cast_fp16")];
            tensor<fp16, [1024, 4096]> blocks_9_mlp_fc2_weight_to_fp16 = const()[name = tensor<string, []>("blocks_9_mlp_fc2_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(288497280)))];
            tensor<fp16, [1, ?, 1024]> linear_41_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = blocks_9_mlp_fc2_weight_to_fp16, x = input_127_cast_fp16)[name = tensor<string, []>("linear_41_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_93_cast_fp16 = add(x = x_91_cast_fp16, y = linear_41_cast_fp16)[name = tensor<string, []>("x_93_cast_fp16")];
            tensor<string, []> x_93_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_93_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_1065_promoted = const()[name = tensor<string, []>("op_1065_promoted"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, ?, 1024]> x_93_cast_fp16_to_fp32 = cast(dtype = x_93_cast_fp16_to_fp32_dtype_0, x = x_93_cast_fp16)[name = tensor<string, []>("cast_118")];
            tensor<fp32, [1, ?, 1024]> var_1075 = pow(x = x_93_cast_fp16_to_fp32, y = var_1065_promoted)[name = tensor<string, []>("op_1075")];
            tensor<int32, [1]> norm_x_41_axes_0 = const()[name = tensor<string, []>("norm_x_41_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> norm_x_41_keep_dims_0 = const()[name = tensor<string, []>("norm_x_41_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, ?, 1]> norm_x_41 = reduce_mean(axes = norm_x_41_axes_0, keep_dims = norm_x_41_keep_dims_0, x = var_1075)[name = tensor<string, []>("norm_x_41")];
            tensor<string, []> norm_x_41_to_fp16_dtype_0 = const()[name = tensor<string, []>("norm_x_41_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_1078_to_fp16 = const()[name = tensor<string, []>("op_1078_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> norm_x_41_to_fp16 = cast(dtype = norm_x_41_to_fp16_dtype_0, x = norm_x_41)[name = tensor<string, []>("cast_117")];
            tensor<fp16, [1, ?, 1]> var_1079_cast_fp16 = add(x = norm_x_41_to_fp16, y = var_1078_to_fp16)[name = tensor<string, []>("op_1079_cast_fp16")];
            tensor<string, []> var_1079_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_1079_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_1080_epsilon_0 = const()[name = tensor<string, []>("op_1080_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, ?, 1]> var_1079_cast_fp16_to_fp32 = cast(dtype = var_1079_cast_fp16_to_fp32_dtype_0, x = var_1079_cast_fp16)[name = tensor<string, []>("cast_116")];
            tensor<fp32, [1, ?, 1]> var_1080 = rsqrt(epsilon = var_1080_epsilon_0, x = var_1079_cast_fp16_to_fp32)[name = tensor<string, []>("op_1080")];
            tensor<string, []> var_1080_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_1080_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, ?, 1]> var_1080_to_fp16 = cast(dtype = var_1080_to_fp16_dtype_0, x = var_1080)[name = tensor<string, []>("cast_115")];
            tensor<fp16, [1, ?, 1024]> var_1081_cast_fp16 = mul(x = x_93_cast_fp16, y = var_1080_to_fp16)[name = tensor<string, []>("op_1081_cast_fp16")];
            tensor<fp16, [1024]> blocks_10_att_norm_weight_to_fp16 = const()[name = tensor<string, []>("blocks_10_att_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(296885952)))];
            tensor<fp16, [1, ?, 1024]> input_129_cast_fp16 = mul(x = var_1081_cast_fp16, y = blocks_10_att_norm_weight_to_fp16)[name = tensor<string, []>("input_129_cast_fp16")];
            tensor<fp16, [3072, 1024]> blocks_10_att_c_attn_weight_to_fp16 = const()[name = tensor<string, []>("blocks_10_att_c_attn_weight_to_fp16"), val = tensor<fp16, [3072, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(296888064)))];
            tensor<fp16, [1, ?, 3072]> linear_42_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = blocks_10_att_c_attn_weight_to_fp16, x = input_129_cast_fp16)[name = tensor<string, []>("linear_42_cast_fp16")];
            tensor<int32, [3]> tile_10 = const()[name = tensor<string, []>("tile_10"), val = tensor<int32, [3]>([1024, 1024, 1024])];
            tensor<int32, []> var_1089_axis_0 = const()[name = tensor<string, []>("op_1089_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, ?, 1024]> var_1089_cast_fp16_0, tensor<fp16, [1, ?, 1024]> var_1089_cast_fp16_1, tensor<fp16, [1, ?, 1024]> var_1089_cast_fp16_2 = split(axis = var_1089_axis_0, split_sizes = tile_10, x = linear_42_cast_fp16)[name = tensor<string, []>("op_1089_cast_fp16")];
            tensor<int32, [4]> var_1093 = const()[name = tensor<string, []>("op_1093"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_1094_cast_fp16 = reshape(shape = var_1093, x = var_1089_cast_fp16_0)[name = tensor<string, []>("op_1094_cast_fp16")];
            tensor<int32, [4]> x_95_perm_0 = const()[name = tensor<string, []>("x_95_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_1096 = const()[name = tensor<string, []>("op_1096"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_1097_cast_fp16 = reshape(shape = var_1096, x = var_1089_cast_fp16_1)[name = tensor<string, []>("op_1097_cast_fp16")];
            tensor<int32, [4]> x_97_perm_0 = const()[name = tensor<string, []>("x_97_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_1099 = const()[name = tensor<string, []>("op_1099"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_1100_cast_fp16 = reshape(shape = var_1099, x = var_1089_cast_fp16_2)[name = tensor<string, []>("op_1100_cast_fp16")];
            tensor<int32, [4]> v_43_perm_0 = const()[name = tensor<string, []>("v_43_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> x1_41_begin_0 = const()[name = tensor<string, []>("x1_41_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_41_end_0 = const()[name = tensor<string, []>("x1_41_end_0"), val = tensor<int32, [4]>([1, 16, 0, 64])];
            tensor<bool, [4]> x1_41_end_mask_0 = const()[name = tensor<string, []>("x1_41_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> x1_41_stride_0 = const()[name = tensor<string, []>("x1_41_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 16, ?, 64]> x_95_cast_fp16 = transpose(perm = x_95_perm_0, x = var_1094_cast_fp16)[name = tensor<string, []>("transpose_10")];
            tensor<fp16, [1, 16, ?, 32]> x1_41_cast_fp16 = slice_by_index(begin = x1_41_begin_0, end = x1_41_end_0, end_mask = x1_41_end_mask_0, stride = x1_41_stride_0, x = x_95_cast_fp16)[name = tensor<string, []>("x1_41_cast_fp16")];
            tensor<int32, [4]> x2_41_begin_0 = const()[name = tensor<string, []>("x2_41_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 1])];
            tensor<int32, [4]> x2_41_end_0 = const()[name = tensor<string, []>("x2_41_end_0"), val = tensor<int32, [4]>([1, 16, 0, 64])];
            tensor<bool, [4]> x2_41_end_mask_0 = const()[name = tensor<string, []>("x2_41_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> x2_41_stride_0 = const()[name = tensor<string, []>("x2_41_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 16, ?, 32]> x2_41_cast_fp16 = slice_by_index(begin = x2_41_begin_0, end = x2_41_end_0, end_mask = x2_41_end_mask_0, stride = x2_41_stride_0, x = x_95_cast_fp16)[name = tensor<string, []>("x2_41_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_1104_cast_fp16 = mul(x = x1_41_cast_fp16, y = blocks_0_att_head_cos_to_fp16)[name = tensor<string, []>("op_1104_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_1105_cast_fp16 = mul(x = x2_41_cast_fp16, y = blocks_0_att_head_sin_to_fp16)[name = tensor<string, []>("op_1105_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_1106_cast_fp16 = sub(x = var_1104_cast_fp16, y = var_1105_cast_fp16)[name = tensor<string, []>("op_1106_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_1107_cast_fp16 = mul(x = x2_41_cast_fp16, y = blocks_0_att_head_cos_to_fp16)[name = tensor<string, []>("op_1107_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_1108_cast_fp16 = mul(x = x1_41_cast_fp16, y = blocks_0_att_head_sin_to_fp16)[name = tensor<string, []>("op_1108_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_1109_cast_fp16 = add(x = var_1107_cast_fp16, y = var_1108_cast_fp16)[name = tensor<string, []>("op_1109_cast_fp16")];
            tensor<int32, []> out_41_axis_0 = const()[name = tensor<string, []>("out_41_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 16, ?, 32, 2]> out_41_cast_fp16 = stack(axis = out_41_axis_0, values = (var_1106_cast_fp16, var_1109_cast_fp16))[name = tensor<string, []>("out_41_cast_fp16")];
            tensor<int32, [4]> concat_24x = const()[name = tensor<string, []>("concat_24x"), val = tensor<int32, [4]>([1, 16, -1, 64])];
            tensor<fp16, [1, 16, ?, 64]> q_43_cast_fp16 = reshape(shape = concat_24x, x = out_41_cast_fp16)[name = tensor<string, []>("q_43_cast_fp16")];
            tensor<int32, [4]> x1_43_begin_0 = const()[name = tensor<string, []>("x1_43_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_43_end_0 = const()[name = tensor<string, []>("x1_43_end_0"), val = tensor<int32, [4]>([1, 16, 0, 64])];
            tensor<bool, [4]> x1_43_end_mask_0 = const()[name = tensor<string, []>("x1_43_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> x1_43_stride_0 = const()[name = tensor<string, []>("x1_43_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 16, ?, 64]> x_97_cast_fp16 = transpose(perm = x_97_perm_0, x = var_1097_cast_fp16)[name = tensor<string, []>("transpose_9")];
            tensor<fp16, [1, 16, ?, 32]> x1_43_cast_fp16 = slice_by_index(begin = x1_43_begin_0, end = x1_43_end_0, end_mask = x1_43_end_mask_0, stride = x1_43_stride_0, x = x_97_cast_fp16)[name = tensor<string, []>("x1_43_cast_fp16")];
            tensor<int32, [4]> x2_43_begin_0 = const()[name = tensor<string, []>("x2_43_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 1])];
            tensor<int32, [4]> x2_43_end_0 = const()[name = tensor<string, []>("x2_43_end_0"), val = tensor<int32, [4]>([1, 16, 0, 64])];
            tensor<bool, [4]> x2_43_end_mask_0 = const()[name = tensor<string, []>("x2_43_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> x2_43_stride_0 = const()[name = tensor<string, []>("x2_43_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 16, ?, 32]> x2_43_cast_fp16 = slice_by_index(begin = x2_43_begin_0, end = x2_43_end_0, end_mask = x2_43_end_mask_0, stride = x2_43_stride_0, x = x_97_cast_fp16)[name = tensor<string, []>("x2_43_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_1115_cast_fp16 = mul(x = x1_43_cast_fp16, y = blocks_0_att_head_cos_to_fp16)[name = tensor<string, []>("op_1115_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_1116_cast_fp16 = mul(x = x2_43_cast_fp16, y = blocks_0_att_head_sin_to_fp16)[name = tensor<string, []>("op_1116_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_1117_cast_fp16 = sub(x = var_1115_cast_fp16, y = var_1116_cast_fp16)[name = tensor<string, []>("op_1117_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_1118_cast_fp16 = mul(x = x2_43_cast_fp16, y = blocks_0_att_head_cos_to_fp16)[name = tensor<string, []>("op_1118_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_1119_cast_fp16 = mul(x = x1_43_cast_fp16, y = blocks_0_att_head_sin_to_fp16)[name = tensor<string, []>("op_1119_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_1120_cast_fp16 = add(x = var_1118_cast_fp16, y = var_1119_cast_fp16)[name = tensor<string, []>("op_1120_cast_fp16")];
            tensor<int32, []> out_43_axis_0 = const()[name = tensor<string, []>("out_43_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 16, ?, 32, 2]> out_43_cast_fp16 = stack(axis = out_43_axis_0, values = (var_1117_cast_fp16, var_1120_cast_fp16))[name = tensor<string, []>("out_43_cast_fp16")];
            tensor<int32, [4]> concat_25x = const()[name = tensor<string, []>("concat_25x"), val = tensor<int32, [4]>([1, 16, -1, 64])];
            tensor<fp16, [1, 16, ?, 64]> k_43_cast_fp16 = reshape(shape = concat_25x, x = out_43_cast_fp16)[name = tensor<string, []>("k_43_cast_fp16")];
            tensor<fp16, []> mul_36_y_0_to_fp16 = const()[name = tensor<string, []>("mul_36_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 16, ?, 64]> mul_36_cast_fp16 = mul(x = q_43_cast_fp16, y = mul_36_y_0_to_fp16)[name = tensor<string, []>("mul_36_cast_fp16")];
            tensor<bool, []> matmul_10_transpose_y_0 = const()[name = tensor<string, []>("matmul_10_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_10_transpose_x_0 = const()[name = tensor<string, []>("matmul_10_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 16, ?, ?]> matmul_10_cast_fp16 = matmul(transpose_x = matmul_10_transpose_x_0, transpose_y = matmul_10_transpose_y_0, x = mul_36_cast_fp16, y = k_43_cast_fp16)[name = tensor<string, []>("matmul_10_cast_fp16")];
            tensor<string, []> matmul_10_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("matmul_10_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<int32, []> softmax_10_axis_0 = const()[name = tensor<string, []>("softmax_10_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp32, [1, 16, ?, ?]> matmul_10_cast_fp16_to_fp32 = cast(dtype = matmul_10_cast_fp16_to_fp32_dtype_0, x = matmul_10_cast_fp16)[name = tensor<string, []>("cast_114")];
            tensor<fp32, [1, 16, ?, ?]> softmax_10 = softmax(axis = softmax_10_axis_0, x = matmul_10_cast_fp16_to_fp32)[name = tensor<string, []>("softmax_10")];
            tensor<bool, []> y_21_transpose_x_0 = const()[name = tensor<string, []>("y_21_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> y_21_transpose_y_0 = const()[name = tensor<string, []>("y_21_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<string, []> softmax_10_to_fp16_dtype_0 = const()[name = tensor<string, []>("softmax_10_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 16, ?, ?]> softmax_10_to_fp16 = cast(dtype = softmax_10_to_fp16_dtype_0, x = softmax_10)[name = tensor<string, []>("cast_113")];
            tensor<fp16, [1, 16, ?, 64]> v_43_cast_fp16 = transpose(perm = v_43_perm_0, x = var_1100_cast_fp16)[name = tensor<string, []>("transpose_8")];
            tensor<fp16, [1, 16, ?, 64]> y_21_cast_fp16 = matmul(transpose_x = y_21_transpose_x_0, transpose_y = y_21_transpose_y_0, x = softmax_10_to_fp16, y = v_43_cast_fp16)[name = tensor<string, []>("y_21_cast_fp16")];
            tensor<int32, [4]> var_1125_perm_0 = const()[name = tensor<string, []>("op_1125_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1126 = const()[name = tensor<string, []>("op_1126"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, ?, 16, 64]> var_1125_cast_fp16 = transpose(perm = var_1125_perm_0, x = y_21_cast_fp16)[name = tensor<string, []>("transpose_7")];
            tensor<fp16, [1, ?, 1024]> input_131_cast_fp16 = reshape(shape = var_1126, x = var_1125_cast_fp16)[name = tensor<string, []>("input_131_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_10_att_c_proj_weight_to_fp16 = const()[name = tensor<string, []>("blocks_10_att_c_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(303179584)))];
            tensor<fp16, [1, ?, 1024]> linear_43_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = blocks_10_att_c_proj_weight_to_fp16, x = input_131_cast_fp16)[name = tensor<string, []>("linear_43_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_99_cast_fp16 = add(x = x_93_cast_fp16, y = linear_43_cast_fp16)[name = tensor<string, []>("x_99_cast_fp16")];
            tensor<string, []> x_99_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_99_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_1065_promoted_1 = const()[name = tensor<string, []>("op_1065_promoted_1"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, ?, 1024]> x_99_cast_fp16_to_fp32 = cast(dtype = x_99_cast_fp16_to_fp32_dtype_0, x = x_99_cast_fp16)[name = tensor<string, []>("cast_112")];
            tensor<fp32, [1, ?, 1024]> var_1132 = pow(x = x_99_cast_fp16_to_fp32, y = var_1065_promoted_1)[name = tensor<string, []>("op_1132")];
            tensor<int32, [1]> norm_x_43_axes_0 = const()[name = tensor<string, []>("norm_x_43_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> norm_x_43_keep_dims_0 = const()[name = tensor<string, []>("norm_x_43_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, ?, 1]> norm_x_43 = reduce_mean(axes = norm_x_43_axes_0, keep_dims = norm_x_43_keep_dims_0, x = var_1132)[name = tensor<string, []>("norm_x_43")];
            tensor<string, []> norm_x_43_to_fp16_dtype_0 = const()[name = tensor<string, []>("norm_x_43_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_1135_to_fp16 = const()[name = tensor<string, []>("op_1135_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> norm_x_43_to_fp16 = cast(dtype = norm_x_43_to_fp16_dtype_0, x = norm_x_43)[name = tensor<string, []>("cast_111")];
            tensor<fp16, [1, ?, 1]> var_1136_cast_fp16 = add(x = norm_x_43_to_fp16, y = var_1135_to_fp16)[name = tensor<string, []>("op_1136_cast_fp16")];
            tensor<string, []> var_1136_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_1136_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_1137_epsilon_0 = const()[name = tensor<string, []>("op_1137_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, ?, 1]> var_1136_cast_fp16_to_fp32 = cast(dtype = var_1136_cast_fp16_to_fp32_dtype_0, x = var_1136_cast_fp16)[name = tensor<string, []>("cast_110")];
            tensor<fp32, [1, ?, 1]> var_1137 = rsqrt(epsilon = var_1137_epsilon_0, x = var_1136_cast_fp16_to_fp32)[name = tensor<string, []>("op_1137")];
            tensor<string, []> var_1137_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_1137_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, ?, 1]> var_1137_to_fp16 = cast(dtype = var_1137_to_fp16_dtype_0, x = var_1137)[name = tensor<string, []>("cast_109")];
            tensor<fp16, [1, ?, 1024]> var_1138_cast_fp16 = mul(x = x_99_cast_fp16, y = var_1137_to_fp16)[name = tensor<string, []>("op_1138_cast_fp16")];
            tensor<fp16, [1024]> blocks_10_ffn_norm_weight_to_fp16 = const()[name = tensor<string, []>("blocks_10_ffn_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(305276800)))];
            tensor<fp16, [1, ?, 1024]> input_133_cast_fp16 = mul(x = var_1138_cast_fp16, y = blocks_10_ffn_norm_weight_to_fp16)[name = tensor<string, []>("input_133_cast_fp16")];
            tensor<fp16, [4096, 1024]> blocks_10_mlp_fc1_weight_to_fp16 = const()[name = tensor<string, []>("blocks_10_mlp_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(305278912)))];
            tensor<fp16, [1, ?, 4096]> linear_44_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_10_mlp_fc1_weight_to_fp16, x = input_133_cast_fp16)[name = tensor<string, []>("linear_44_cast_fp16")];
            tensor<fp16, [1, ?, 4096]> input_137_cast_fp16 = silu(x = linear_44_cast_fp16)[name = tensor<string, []>("input_137_cast_fp16")];
            tensor<fp16, [1024, 4096]> blocks_10_mlp_fc2_weight_to_fp16 = const()[name = tensor<string, []>("blocks_10_mlp_fc2_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(313667584)))];
            tensor<fp16, [1, ?, 1024]> linear_45_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = blocks_10_mlp_fc2_weight_to_fp16, x = input_137_cast_fp16)[name = tensor<string, []>("linear_45_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_101_cast_fp16 = add(x = x_99_cast_fp16, y = linear_45_cast_fp16)[name = tensor<string, []>("x_101_cast_fp16")];
            tensor<string, []> x_101_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_101_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_1157_promoted = const()[name = tensor<string, []>("op_1157_promoted"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, ?, 1024]> x_101_cast_fp16_to_fp32 = cast(dtype = x_101_cast_fp16_to_fp32_dtype_0, x = x_101_cast_fp16)[name = tensor<string, []>("cast_108")];
            tensor<fp32, [1, ?, 1024]> var_1167 = pow(x = x_101_cast_fp16_to_fp32, y = var_1157_promoted)[name = tensor<string, []>("op_1167")];
            tensor<int32, [1]> norm_x_45_axes_0 = const()[name = tensor<string, []>("norm_x_45_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> norm_x_45_keep_dims_0 = const()[name = tensor<string, []>("norm_x_45_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, ?, 1]> norm_x_45 = reduce_mean(axes = norm_x_45_axes_0, keep_dims = norm_x_45_keep_dims_0, x = var_1167)[name = tensor<string, []>("norm_x_45")];
            tensor<string, []> norm_x_45_to_fp16_dtype_0 = const()[name = tensor<string, []>("norm_x_45_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_1170_to_fp16 = const()[name = tensor<string, []>("op_1170_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> norm_x_45_to_fp16 = cast(dtype = norm_x_45_to_fp16_dtype_0, x = norm_x_45)[name = tensor<string, []>("cast_107")];
            tensor<fp16, [1, ?, 1]> var_1171_cast_fp16 = add(x = norm_x_45_to_fp16, y = var_1170_to_fp16)[name = tensor<string, []>("op_1171_cast_fp16")];
            tensor<string, []> var_1171_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_1171_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_1172_epsilon_0 = const()[name = tensor<string, []>("op_1172_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, ?, 1]> var_1171_cast_fp16_to_fp32 = cast(dtype = var_1171_cast_fp16_to_fp32_dtype_0, x = var_1171_cast_fp16)[name = tensor<string, []>("cast_106")];
            tensor<fp32, [1, ?, 1]> var_1172 = rsqrt(epsilon = var_1172_epsilon_0, x = var_1171_cast_fp16_to_fp32)[name = tensor<string, []>("op_1172")];
            tensor<string, []> var_1172_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_1172_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, ?, 1]> var_1172_to_fp16 = cast(dtype = var_1172_to_fp16_dtype_0, x = var_1172)[name = tensor<string, []>("cast_105")];
            tensor<fp16, [1, ?, 1024]> var_1173_cast_fp16 = mul(x = x_101_cast_fp16, y = var_1172_to_fp16)[name = tensor<string, []>("op_1173_cast_fp16")];
            tensor<fp16, [1024]> blocks_11_att_norm_weight_to_fp16 = const()[name = tensor<string, []>("blocks_11_att_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(322056256)))];
            tensor<fp16, [1, ?, 1024]> input_139_cast_fp16 = mul(x = var_1173_cast_fp16, y = blocks_11_att_norm_weight_to_fp16)[name = tensor<string, []>("input_139_cast_fp16")];
            tensor<fp16, [3072, 1024]> blocks_11_att_c_attn_weight_to_fp16 = const()[name = tensor<string, []>("blocks_11_att_c_attn_weight_to_fp16"), val = tensor<fp16, [3072, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(322058368)))];
            tensor<fp16, [1, ?, 3072]> linear_46_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = blocks_11_att_c_attn_weight_to_fp16, x = input_139_cast_fp16)[name = tensor<string, []>("linear_46_cast_fp16")];
            tensor<int32, [3]> tile_11 = const()[name = tensor<string, []>("tile_11"), val = tensor<int32, [3]>([1024, 1024, 1024])];
            tensor<int32, []> var_1181_axis_0 = const()[name = tensor<string, []>("op_1181_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, ?, 1024]> var_1181_cast_fp16_0, tensor<fp16, [1, ?, 1024]> var_1181_cast_fp16_1, tensor<fp16, [1, ?, 1024]> var_1181_cast_fp16_2 = split(axis = var_1181_axis_0, split_sizes = tile_11, x = linear_46_cast_fp16)[name = tensor<string, []>("op_1181_cast_fp16")];
            tensor<int32, [4]> var_1185 = const()[name = tensor<string, []>("op_1185"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_1186_cast_fp16 = reshape(shape = var_1185, x = var_1181_cast_fp16_0)[name = tensor<string, []>("op_1186_cast_fp16")];
            tensor<int32, [4]> x_103_perm_0 = const()[name = tensor<string, []>("x_103_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_1188 = const()[name = tensor<string, []>("op_1188"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_1189_cast_fp16 = reshape(shape = var_1188, x = var_1181_cast_fp16_1)[name = tensor<string, []>("op_1189_cast_fp16")];
            tensor<int32, [4]> x_105_perm_0 = const()[name = tensor<string, []>("x_105_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_1191 = const()[name = tensor<string, []>("op_1191"), val = tensor<int32, [4]>([1, -1, 16, 64])];
            tensor<fp16, [1, ?, 16, 64]> var_1192_cast_fp16 = reshape(shape = var_1191, x = var_1181_cast_fp16_2)[name = tensor<string, []>("op_1192_cast_fp16")];
            tensor<int32, [4]> v_perm_0 = const()[name = tensor<string, []>("v_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> x1_45_begin_0 = const()[name = tensor<string, []>("x1_45_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_45_end_0 = const()[name = tensor<string, []>("x1_45_end_0"), val = tensor<int32, [4]>([1, 16, 0, 64])];
            tensor<bool, [4]> x1_45_end_mask_0 = const()[name = tensor<string, []>("x1_45_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> x1_45_stride_0 = const()[name = tensor<string, []>("x1_45_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 16, ?, 64]> x_103_cast_fp16 = transpose(perm = x_103_perm_0, x = var_1186_cast_fp16)[name = tensor<string, []>("transpose_6")];
            tensor<fp16, [1, 16, ?, 32]> x1_45_cast_fp16 = slice_by_index(begin = x1_45_begin_0, end = x1_45_end_0, end_mask = x1_45_end_mask_0, stride = x1_45_stride_0, x = x_103_cast_fp16)[name = tensor<string, []>("x1_45_cast_fp16")];
            tensor<int32, [4]> x2_45_begin_0 = const()[name = tensor<string, []>("x2_45_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 1])];
            tensor<int32, [4]> x2_45_end_0 = const()[name = tensor<string, []>("x2_45_end_0"), val = tensor<int32, [4]>([1, 16, 0, 64])];
            tensor<bool, [4]> x2_45_end_mask_0 = const()[name = tensor<string, []>("x2_45_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> x2_45_stride_0 = const()[name = tensor<string, []>("x2_45_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 16, ?, 32]> x2_45_cast_fp16 = slice_by_index(begin = x2_45_begin_0, end = x2_45_end_0, end_mask = x2_45_end_mask_0, stride = x2_45_stride_0, x = x_103_cast_fp16)[name = tensor<string, []>("x2_45_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_1196_cast_fp16 = mul(x = x1_45_cast_fp16, y = blocks_0_att_head_cos_to_fp16)[name = tensor<string, []>("op_1196_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_1197_cast_fp16 = mul(x = x2_45_cast_fp16, y = blocks_0_att_head_sin_to_fp16)[name = tensor<string, []>("op_1197_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_1198_cast_fp16 = sub(x = var_1196_cast_fp16, y = var_1197_cast_fp16)[name = tensor<string, []>("op_1198_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_1199_cast_fp16 = mul(x = x2_45_cast_fp16, y = blocks_0_att_head_cos_to_fp16)[name = tensor<string, []>("op_1199_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_1200_cast_fp16 = mul(x = x1_45_cast_fp16, y = blocks_0_att_head_sin_to_fp16)[name = tensor<string, []>("op_1200_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_1201_cast_fp16 = add(x = var_1199_cast_fp16, y = var_1200_cast_fp16)[name = tensor<string, []>("op_1201_cast_fp16")];
            tensor<int32, []> out_45_axis_0 = const()[name = tensor<string, []>("out_45_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 16, ?, 32, 2]> out_45_cast_fp16 = stack(axis = out_45_axis_0, values = (var_1198_cast_fp16, var_1201_cast_fp16))[name = tensor<string, []>("out_45_cast_fp16")];
            tensor<int32, [4]> concat_26x = const()[name = tensor<string, []>("concat_26x"), val = tensor<int32, [4]>([1, 16, -1, 64])];
            tensor<fp16, [1, 16, ?, 64]> q_cast_fp16 = reshape(shape = concat_26x, x = out_45_cast_fp16)[name = tensor<string, []>("q_cast_fp16")];
            tensor<int32, [4]> x1_begin_0 = const()[name = tensor<string, []>("x1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_end_0 = const()[name = tensor<string, []>("x1_end_0"), val = tensor<int32, [4]>([1, 16, 0, 64])];
            tensor<bool, [4]> x1_end_mask_0 = const()[name = tensor<string, []>("x1_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> x1_stride_0 = const()[name = tensor<string, []>("x1_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 16, ?, 64]> x_105_cast_fp16 = transpose(perm = x_105_perm_0, x = var_1189_cast_fp16)[name = tensor<string, []>("transpose_5")];
            tensor<fp16, [1, 16, ?, 32]> x1_cast_fp16 = slice_by_index(begin = x1_begin_0, end = x1_end_0, end_mask = x1_end_mask_0, stride = x1_stride_0, x = x_105_cast_fp16)[name = tensor<string, []>("x1_cast_fp16")];
            tensor<int32, [4]> x2_begin_0 = const()[name = tensor<string, []>("x2_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 1])];
            tensor<int32, [4]> x2_end_0 = const()[name = tensor<string, []>("x2_end_0"), val = tensor<int32, [4]>([1, 16, 0, 64])];
            tensor<bool, [4]> x2_end_mask_0 = const()[name = tensor<string, []>("x2_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> x2_stride_0 = const()[name = tensor<string, []>("x2_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 16, ?, 32]> x2_cast_fp16 = slice_by_index(begin = x2_begin_0, end = x2_end_0, end_mask = x2_end_mask_0, stride = x2_stride_0, x = x_105_cast_fp16)[name = tensor<string, []>("x2_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_1207_cast_fp16 = mul(x = x1_cast_fp16, y = blocks_0_att_head_cos_to_fp16)[name = tensor<string, []>("op_1207_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_1208_cast_fp16 = mul(x = x2_cast_fp16, y = blocks_0_att_head_sin_to_fp16)[name = tensor<string, []>("op_1208_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_1209_cast_fp16 = sub(x = var_1207_cast_fp16, y = var_1208_cast_fp16)[name = tensor<string, []>("op_1209_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_1210_cast_fp16 = mul(x = x2_cast_fp16, y = blocks_0_att_head_cos_to_fp16)[name = tensor<string, []>("op_1210_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_1211_cast_fp16 = mul(x = x1_cast_fp16, y = blocks_0_att_head_sin_to_fp16)[name = tensor<string, []>("op_1211_cast_fp16")];
            tensor<fp16, [1, 16, ?, 32]> var_1212_cast_fp16 = add(x = var_1210_cast_fp16, y = var_1211_cast_fp16)[name = tensor<string, []>("op_1212_cast_fp16")];
            tensor<int32, []> out_axis_0 = const()[name = tensor<string, []>("out_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp16, [1, 16, ?, 32, 2]> out_cast_fp16 = stack(axis = out_axis_0, values = (var_1209_cast_fp16, var_1212_cast_fp16))[name = tensor<string, []>("out_cast_fp16")];
            tensor<int32, [4]> concat_27x = const()[name = tensor<string, []>("concat_27x"), val = tensor<int32, [4]>([1, 16, -1, 64])];
            tensor<fp16, [1, 16, ?, 64]> k_cast_fp16 = reshape(shape = concat_27x, x = out_cast_fp16)[name = tensor<string, []>("k_cast_fp16")];
            tensor<fp16, []> mul_39_y_0_to_fp16 = const()[name = tensor<string, []>("mul_39_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
            tensor<fp16, [1, 16, ?, 64]> mul_39_cast_fp16 = mul(x = q_cast_fp16, y = mul_39_y_0_to_fp16)[name = tensor<string, []>("mul_39_cast_fp16")];
            tensor<bool, []> matmul_11_transpose_y_0 = const()[name = tensor<string, []>("matmul_11_transpose_y_0"), val = tensor<bool, []>(true)];
            tensor<bool, []> matmul_11_transpose_x_0 = const()[name = tensor<string, []>("matmul_11_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 16, ?, ?]> matmul_11_cast_fp16 = matmul(transpose_x = matmul_11_transpose_x_0, transpose_y = matmul_11_transpose_y_0, x = mul_39_cast_fp16, y = k_cast_fp16)[name = tensor<string, []>("matmul_11_cast_fp16")];
            tensor<string, []> matmul_11_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("matmul_11_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<int32, []> softmax_11_axis_0 = const()[name = tensor<string, []>("softmax_11_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp32, [1, 16, ?, ?]> matmul_11_cast_fp16_to_fp32 = cast(dtype = matmul_11_cast_fp16_to_fp32_dtype_0, x = matmul_11_cast_fp16)[name = tensor<string, []>("cast_104")];
            tensor<fp32, [1, 16, ?, ?]> softmax_11 = softmax(axis = softmax_11_axis_0, x = matmul_11_cast_fp16_to_fp32)[name = tensor<string, []>("softmax_11")];
            tensor<bool, []> y_transpose_x_0 = const()[name = tensor<string, []>("y_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> y_transpose_y_0 = const()[name = tensor<string, []>("y_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<string, []> softmax_11_to_fp16_dtype_0 = const()[name = tensor<string, []>("softmax_11_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 16, ?, ?]> softmax_11_to_fp16 = cast(dtype = softmax_11_to_fp16_dtype_0, x = softmax_11)[name = tensor<string, []>("cast_103")];
            tensor<fp16, [1, 16, ?, 64]> v_cast_fp16 = transpose(perm = v_perm_0, x = var_1192_cast_fp16)[name = tensor<string, []>("transpose_4")];
            tensor<fp16, [1, 16, ?, 64]> y_cast_fp16 = matmul(transpose_x = y_transpose_x_0, transpose_y = y_transpose_y_0, x = softmax_11_to_fp16, y = v_cast_fp16)[name = tensor<string, []>("y_cast_fp16")];
            tensor<int32, [4]> var_1217_perm_0 = const()[name = tensor<string, []>("op_1217_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1218 = const()[name = tensor<string, []>("op_1218"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, ?, 16, 64]> var_1217_cast_fp16 = transpose(perm = var_1217_perm_0, x = y_cast_fp16)[name = tensor<string, []>("transpose_3")];
            tensor<fp16, [1, ?, 1024]> input_141_cast_fp16 = reshape(shape = var_1218, x = var_1217_cast_fp16)[name = tensor<string, []>("input_141_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_11_att_c_proj_weight_to_fp16 = const()[name = tensor<string, []>("blocks_11_att_c_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(328349888)))];
            tensor<fp16, [1, ?, 1024]> linear_47_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = blocks_11_att_c_proj_weight_to_fp16, x = input_141_cast_fp16)[name = tensor<string, []>("linear_47_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_107_cast_fp16 = add(x = x_101_cast_fp16, y = linear_47_cast_fp16)[name = tensor<string, []>("x_107_cast_fp16")];
            tensor<string, []> x_107_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_107_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_1157_promoted_1 = const()[name = tensor<string, []>("op_1157_promoted_1"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, ?, 1024]> x_107_cast_fp16_to_fp32 = cast(dtype = x_107_cast_fp16_to_fp32_dtype_0, x = x_107_cast_fp16)[name = tensor<string, []>("cast_102")];
            tensor<fp32, [1, ?, 1024]> var_1224 = pow(x = x_107_cast_fp16_to_fp32, y = var_1157_promoted_1)[name = tensor<string, []>("op_1224")];
            tensor<int32, [1]> norm_x_axes_0 = const()[name = tensor<string, []>("norm_x_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> norm_x_keep_dims_0 = const()[name = tensor<string, []>("norm_x_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, ?, 1]> norm_x = reduce_mean(axes = norm_x_axes_0, keep_dims = norm_x_keep_dims_0, x = var_1224)[name = tensor<string, []>("norm_x")];
            tensor<string, []> norm_x_to_fp16_dtype_0 = const()[name = tensor<string, []>("norm_x_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_1227_to_fp16 = const()[name = tensor<string, []>("op_1227_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> norm_x_to_fp16 = cast(dtype = norm_x_to_fp16_dtype_0, x = norm_x)[name = tensor<string, []>("cast_101")];
            tensor<fp16, [1, ?, 1]> var_1228_cast_fp16 = add(x = norm_x_to_fp16, y = var_1227_to_fp16)[name = tensor<string, []>("op_1228_cast_fp16")];
            tensor<string, []> var_1228_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_1228_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_1229_epsilon_0 = const()[name = tensor<string, []>("op_1229_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, ?, 1]> var_1228_cast_fp16_to_fp32 = cast(dtype = var_1228_cast_fp16_to_fp32_dtype_0, x = var_1228_cast_fp16)[name = tensor<string, []>("cast_100")];
            tensor<fp32, [1, ?, 1]> var_1229 = rsqrt(epsilon = var_1229_epsilon_0, x = var_1228_cast_fp16_to_fp32)[name = tensor<string, []>("op_1229")];
            tensor<string, []> var_1229_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_1229_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, ?, 1]> var_1229_to_fp16 = cast(dtype = var_1229_to_fp16_dtype_0, x = var_1229)[name = tensor<string, []>("cast_99")];
            tensor<fp16, [1, ?, 1024]> var_1230_cast_fp16 = mul(x = x_107_cast_fp16, y = var_1229_to_fp16)[name = tensor<string, []>("op_1230_cast_fp16")];
            tensor<fp16, [1024]> blocks_11_ffn_norm_weight_to_fp16 = const()[name = tensor<string, []>("blocks_11_ffn_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(330447104)))];
            tensor<fp16, [1, ?, 1024]> input_143_cast_fp16 = mul(x = var_1230_cast_fp16, y = blocks_11_ffn_norm_weight_to_fp16)[name = tensor<string, []>("input_143_cast_fp16")];
            tensor<fp16, [4096, 1024]> blocks_11_mlp_fc1_weight_to_fp16 = const()[name = tensor<string, []>("blocks_11_mlp_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(330449216)))];
            tensor<fp16, [1, ?, 4096]> linear_48_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_11_mlp_fc1_weight_to_fp16, x = input_143_cast_fp16)[name = tensor<string, []>("linear_48_cast_fp16")];
            tensor<fp16, [1, ?, 4096]> input_147_cast_fp16 = silu(x = linear_48_cast_fp16)[name = tensor<string, []>("input_147_cast_fp16")];
            tensor<fp16, [1024, 4096]> blocks_11_mlp_fc2_weight_to_fp16 = const()[name = tensor<string, []>("blocks_11_mlp_fc2_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(338837888)))];
            tensor<fp16, [1, ?, 1024]> linear_49_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = blocks_11_mlp_fc2_weight_to_fp16, x = input_147_cast_fp16)[name = tensor<string, []>("linear_49_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_109_cast_fp16 = add(x = x_107_cast_fp16, y = linear_49_cast_fp16)[name = tensor<string, []>("x_109_cast_fp16")];
            tensor<int32, [3]> input_149_perm_0 = const()[name = tensor<string, []>("input_149_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 1024, ?]> input_149_cast_fp16 = transpose(perm = input_149_perm_0, x = x_109_cast_fp16)[name = tensor<string, []>("transpose_2")];
            tensor<int32, [3]> shape_40_cast_fp16 = shape(x = input_149_cast_fp16)[name = tensor<string, []>("shape_40_cast_fp16")];
            tensor<int32, [4]> concat_28x = const()[name = tensor<string, []>("concat_28x"), val = tensor<int32, [4]>([1, 32, 32, -1])];
            tensor<fp16, [1, 32, 32, ?]> reshape_16_cast_fp16 = reshape(shape = concat_28x, x = input_149_cast_fp16)[name = tensor<string, []>("reshape_16_cast_fp16")];
            tensor<string, []> reshape_16_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("reshape_16_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<int32, [2]> reduce_mean_12_axes_0 = const()[name = tensor<string, []>("reduce_mean_12_axes_0"), val = tensor<int32, [2]>([2, 3])];
            tensor<bool, []> reduce_mean_12_keep_dims_0 = const()[name = tensor<string, []>("reduce_mean_12_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 32, 32, ?]> reshape_16_cast_fp16_to_fp32 = cast(dtype = reshape_16_cast_fp16_to_fp32_dtype_0, x = reshape_16_cast_fp16)[name = tensor<string, []>("cast_98")];
            tensor<fp32, [1, 32, 1, 1]> reduce_mean_12 = reduce_mean(axes = reduce_mean_12_axes_0, keep_dims = reduce_mean_12_keep_dims_0, x = reshape_16_cast_fp16_to_fp32)[name = tensor<string, []>("reduce_mean_12")];
            tensor<string, []> reduce_mean_12_to_fp16_dtype_0 = const()[name = tensor<string, []>("reduce_mean_12_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 32, 1, 1]> reduce_mean_12_to_fp16 = cast(dtype = reduce_mean_12_to_fp16_dtype_0, x = reduce_mean_12)[name = tensor<string, []>("cast_97")];
            tensor<fp16, [1, 32, 32, ?]> sub_8_cast_fp16 = sub(x = reshape_16_cast_fp16, y = reduce_mean_12_to_fp16)[name = tensor<string, []>("sub_8_cast_fp16")];
            tensor<fp16, [1, 32, 32, ?]> square_4_cast_fp16 = square(x = sub_8_cast_fp16)[name = tensor<string, []>("square_4_cast_fp16")];
            tensor<string, []> square_4_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("square_4_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<int32, [2]> reduce_mean_14_axes_0 = const()[name = tensor<string, []>("reduce_mean_14_axes_0"), val = tensor<int32, [2]>([2, 3])];
            tensor<bool, []> reduce_mean_14_keep_dims_0 = const()[name = tensor<string, []>("reduce_mean_14_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 32, 32, ?]> square_4_cast_fp16_to_fp32 = cast(dtype = square_4_cast_fp16_to_fp32_dtype_0, x = square_4_cast_fp16)[name = tensor<string, []>("cast_96")];
            tensor<fp32, [1, 32, 1, 1]> reduce_mean_14 = reduce_mean(axes = reduce_mean_14_axes_0, keep_dims = reduce_mean_14_keep_dims_0, x = square_4_cast_fp16_to_fp32)[name = tensor<string, []>("reduce_mean_14")];
            tensor<string, []> reduce_mean_14_to_fp16_dtype_0 = const()[name = tensor<string, []>("reduce_mean_14_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> add_8_y_0_to_fp16 = const()[name = tensor<string, []>("add_8_y_0_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 32, 1, 1]> reduce_mean_14_to_fp16 = cast(dtype = reduce_mean_14_to_fp16_dtype_0, x = reduce_mean_14)[name = tensor<string, []>("cast_95")];
            tensor<fp16, [1, 32, 1, 1]> add_8_cast_fp16 = add(x = reduce_mean_14_to_fp16, y = add_8_y_0_to_fp16)[name = tensor<string, []>("add_8_cast_fp16")];
            tensor<fp16, [1, 32, 1, 1]> sqrt_4_cast_fp16 = sqrt(x = add_8_cast_fp16)[name = tensor<string, []>("sqrt_4_cast_fp16")];
            tensor<fp16, [1, 32, 32, ?]> real_div_4_cast_fp16 = real_div(x = sub_8_cast_fp16, y = sqrt_4_cast_fp16)[name = tensor<string, []>("real_div_4_cast_fp16")];
            tensor<fp16, [1, 1024, ?]> reshape_17_cast_fp16 = reshape(shape = shape_40_cast_fp16, x = real_div_4_cast_fp16)[name = tensor<string, []>("reshape_17_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> reshape_18_to_fp16 = const()[name = tensor<string, []>("reshape_18_to_fp16"), val = tensor<fp16, [1, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(347226560)))];
            tensor<fp16, [1, 1024, ?]> mul_40_cast_fp16 = mul(x = reshape_17_cast_fp16, y = reshape_18_to_fp16)[name = tensor<string, []>("mul_40_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> reshape_19_to_fp16 = const()[name = tensor<string, []>("reshape_19_to_fp16"), val = tensor<fp16, [1, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(347228672)))];
            tensor<fp16, [1, 1024, ?]> add_9_cast_fp16 = add(x = mul_40_cast_fp16, y = reshape_19_to_fp16)[name = tensor<string, []>("add_9_cast_fp16")];
            tensor<fp16, [1, 1024, ?]> input_151_cast_fp16 = silu(x = add_9_cast_fp16)[name = tensor<string, []>("input_151_cast_fp16")];
            tensor<string, []> input_153_pad_type_0 = const()[name = tensor<string, []>("input_153_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [2]> input_153_pad_0 = const()[name = tensor<string, []>("input_153_pad_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [1]> input_153_strides_0 = const()[name = tensor<string, []>("input_153_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> input_153_dilations_0 = const()[name = tensor<string, []>("input_153_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> input_153_groups_0 = const()[name = tensor<string, []>("input_153_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 3]> post_net_0_conv1_weight_to_fp16 = const()[name = tensor<string, []>("post_net_0_conv1_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 3]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(347230784)))];
            tensor<fp16, [1024]> post_net_0_conv1_bias_to_fp16 = const()[name = tensor<string, []>("post_net_0_conv1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(353522304)))];
            tensor<fp16, [1, 1024, ?]> input_153_cast_fp16 = conv(bias = post_net_0_conv1_bias_to_fp16, dilations = input_153_dilations_0, groups = input_153_groups_0, pad = input_153_pad_0, pad_type = input_153_pad_type_0, strides = input_153_strides_0, weight = post_net_0_conv1_weight_to_fp16, x = input_151_cast_fp16)[name = tensor<string, []>("input_153_cast_fp16")];
            tensor<int32, [3]> shape_41_cast_fp16 = shape(x = input_153_cast_fp16)[name = tensor<string, []>("shape_41_cast_fp16")];
            tensor<int32, [4]> concat_29x = const()[name = tensor<string, []>("concat_29x"), val = tensor<int32, [4]>([1, 32, 32, -1])];
            tensor<fp16, [1, 32, 32, ?]> reshape_20_cast_fp16 = reshape(shape = concat_29x, x = input_153_cast_fp16)[name = tensor<string, []>("reshape_20_cast_fp16")];
            tensor<string, []> reshape_20_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("reshape_20_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<int32, [2]> reduce_mean_15_axes_0 = const()[name = tensor<string, []>("reduce_mean_15_axes_0"), val = tensor<int32, [2]>([2, 3])];
            tensor<bool, []> reduce_mean_15_keep_dims_0 = const()[name = tensor<string, []>("reduce_mean_15_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 32, 32, ?]> reshape_20_cast_fp16_to_fp32 = cast(dtype = reshape_20_cast_fp16_to_fp32_dtype_0, x = reshape_20_cast_fp16)[name = tensor<string, []>("cast_94")];
            tensor<fp32, [1, 32, 1, 1]> reduce_mean_15 = reduce_mean(axes = reduce_mean_15_axes_0, keep_dims = reduce_mean_15_keep_dims_0, x = reshape_20_cast_fp16_to_fp32)[name = tensor<string, []>("reduce_mean_15")];
            tensor<string, []> reduce_mean_15_to_fp16_dtype_0 = const()[name = tensor<string, []>("reduce_mean_15_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 32, 1, 1]> reduce_mean_15_to_fp16 = cast(dtype = reduce_mean_15_to_fp16_dtype_0, x = reduce_mean_15)[name = tensor<string, []>("cast_93")];
            tensor<fp16, [1, 32, 32, ?]> sub_10_cast_fp16 = sub(x = reshape_20_cast_fp16, y = reduce_mean_15_to_fp16)[name = tensor<string, []>("sub_10_cast_fp16")];
            tensor<fp16, [1, 32, 32, ?]> square_5_cast_fp16 = square(x = sub_10_cast_fp16)[name = tensor<string, []>("square_5_cast_fp16")];
            tensor<string, []> square_5_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("square_5_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<int32, [2]> reduce_mean_17_axes_0 = const()[name = tensor<string, []>("reduce_mean_17_axes_0"), val = tensor<int32, [2]>([2, 3])];
            tensor<bool, []> reduce_mean_17_keep_dims_0 = const()[name = tensor<string, []>("reduce_mean_17_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 32, 32, ?]> square_5_cast_fp16_to_fp32 = cast(dtype = square_5_cast_fp16_to_fp32_dtype_0, x = square_5_cast_fp16)[name = tensor<string, []>("cast_92")];
            tensor<fp32, [1, 32, 1, 1]> reduce_mean_17 = reduce_mean(axes = reduce_mean_17_axes_0, keep_dims = reduce_mean_17_keep_dims_0, x = square_5_cast_fp16_to_fp32)[name = tensor<string, []>("reduce_mean_17")];
            tensor<string, []> reduce_mean_17_to_fp16_dtype_0 = const()[name = tensor<string, []>("reduce_mean_17_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> add_10_y_0_to_fp16 = const()[name = tensor<string, []>("add_10_y_0_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 32, 1, 1]> reduce_mean_17_to_fp16 = cast(dtype = reduce_mean_17_to_fp16_dtype_0, x = reduce_mean_17)[name = tensor<string, []>("cast_91")];
            tensor<fp16, [1, 32, 1, 1]> add_10_cast_fp16 = add(x = reduce_mean_17_to_fp16, y = add_10_y_0_to_fp16)[name = tensor<string, []>("add_10_cast_fp16")];
            tensor<fp16, [1, 32, 1, 1]> sqrt_5_cast_fp16 = sqrt(x = add_10_cast_fp16)[name = tensor<string, []>("sqrt_5_cast_fp16")];
            tensor<fp16, [1, 32, 32, ?]> real_div_5_cast_fp16 = real_div(x = sub_10_cast_fp16, y = sqrt_5_cast_fp16)[name = tensor<string, []>("real_div_5_cast_fp16")];
            tensor<fp16, [1, 1024, ?]> reshape_21_cast_fp16 = reshape(shape = shape_41_cast_fp16, x = real_div_5_cast_fp16)[name = tensor<string, []>("reshape_21_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> reshape_22_to_fp16 = const()[name = tensor<string, []>("reshape_22_to_fp16"), val = tensor<fp16, [1, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(353524416)))];
            tensor<fp16, [1, 1024, ?]> mul_41_cast_fp16 = mul(x = reshape_21_cast_fp16, y = reshape_22_to_fp16)[name = tensor<string, []>("mul_41_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> reshape_23_to_fp16 = const()[name = tensor<string, []>("reshape_23_to_fp16"), val = tensor<fp16, [1, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(353526528)))];
            tensor<fp16, [1, 1024, ?]> add_11_cast_fp16 = add(x = mul_41_cast_fp16, y = reshape_23_to_fp16)[name = tensor<string, []>("add_11_cast_fp16")];
            tensor<fp16, [1, 1024, ?]> input_155_cast_fp16 = silu(x = add_11_cast_fp16)[name = tensor<string, []>("input_155_cast_fp16")];
            tensor<string, []> h_5_pad_type_0 = const()[name = tensor<string, []>("h_5_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [2]> h_5_pad_0 = const()[name = tensor<string, []>("h_5_pad_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [1]> h_5_strides_0 = const()[name = tensor<string, []>("h_5_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> h_5_dilations_0 = const()[name = tensor<string, []>("h_5_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> h_5_groups_0 = const()[name = tensor<string, []>("h_5_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 3]> post_net_0_conv2_weight_to_fp16 = const()[name = tensor<string, []>("post_net_0_conv2_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 3]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(353528640)))];
            tensor<fp16, [1024]> post_net_0_conv2_bias_to_fp16 = const()[name = tensor<string, []>("post_net_0_conv2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(359820160)))];
            tensor<fp16, [1, 1024, ?]> h_5_cast_fp16 = conv(bias = post_net_0_conv2_bias_to_fp16, dilations = h_5_dilations_0, groups = h_5_groups_0, pad = h_5_pad_0, pad_type = h_5_pad_type_0, strides = h_5_strides_0, weight = post_net_0_conv2_weight_to_fp16, x = input_155_cast_fp16)[name = tensor<string, []>("h_5_cast_fp16")];
            tensor<fp16, [1, 1024, ?]> input_159_cast_fp16 = add(x = input_149_cast_fp16, y = h_5_cast_fp16)[name = tensor<string, []>("input_159_cast_fp16")];
            tensor<int32, [3]> shape_42_cast_fp16 = shape(x = input_159_cast_fp16)[name = tensor<string, []>("shape_42_cast_fp16")];
            tensor<int32, [4]> concat_30x = const()[name = tensor<string, []>("concat_30x"), val = tensor<int32, [4]>([1, 32, 32, -1])];
            tensor<fp16, [1, 32, 32, ?]> reshape_24_cast_fp16 = reshape(shape = concat_30x, x = input_159_cast_fp16)[name = tensor<string, []>("reshape_24_cast_fp16")];
            tensor<string, []> reshape_24_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("reshape_24_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<int32, [2]> reduce_mean_18_axes_0 = const()[name = tensor<string, []>("reduce_mean_18_axes_0"), val = tensor<int32, [2]>([2, 3])];
            tensor<bool, []> reduce_mean_18_keep_dims_0 = const()[name = tensor<string, []>("reduce_mean_18_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 32, 32, ?]> reshape_24_cast_fp16_to_fp32 = cast(dtype = reshape_24_cast_fp16_to_fp32_dtype_0, x = reshape_24_cast_fp16)[name = tensor<string, []>("cast_90")];
            tensor<fp32, [1, 32, 1, 1]> reduce_mean_18 = reduce_mean(axes = reduce_mean_18_axes_0, keep_dims = reduce_mean_18_keep_dims_0, x = reshape_24_cast_fp16_to_fp32)[name = tensor<string, []>("reduce_mean_18")];
            tensor<string, []> reduce_mean_18_to_fp16_dtype_0 = const()[name = tensor<string, []>("reduce_mean_18_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 32, 1, 1]> reduce_mean_18_to_fp16 = cast(dtype = reduce_mean_18_to_fp16_dtype_0, x = reduce_mean_18)[name = tensor<string, []>("cast_89")];
            tensor<fp16, [1, 32, 32, ?]> sub_12_cast_fp16 = sub(x = reshape_24_cast_fp16, y = reduce_mean_18_to_fp16)[name = tensor<string, []>("sub_12_cast_fp16")];
            tensor<fp16, [1, 32, 32, ?]> square_6_cast_fp16 = square(x = sub_12_cast_fp16)[name = tensor<string, []>("square_6_cast_fp16")];
            tensor<string, []> square_6_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("square_6_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<int32, [2]> reduce_mean_20_axes_0 = const()[name = tensor<string, []>("reduce_mean_20_axes_0"), val = tensor<int32, [2]>([2, 3])];
            tensor<bool, []> reduce_mean_20_keep_dims_0 = const()[name = tensor<string, []>("reduce_mean_20_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 32, 32, ?]> square_6_cast_fp16_to_fp32 = cast(dtype = square_6_cast_fp16_to_fp32_dtype_0, x = square_6_cast_fp16)[name = tensor<string, []>("cast_88")];
            tensor<fp32, [1, 32, 1, 1]> reduce_mean_20 = reduce_mean(axes = reduce_mean_20_axes_0, keep_dims = reduce_mean_20_keep_dims_0, x = square_6_cast_fp16_to_fp32)[name = tensor<string, []>("reduce_mean_20")];
            tensor<string, []> reduce_mean_20_to_fp16_dtype_0 = const()[name = tensor<string, []>("reduce_mean_20_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> add_12_y_0_to_fp16 = const()[name = tensor<string, []>("add_12_y_0_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 32, 1, 1]> reduce_mean_20_to_fp16 = cast(dtype = reduce_mean_20_to_fp16_dtype_0, x = reduce_mean_20)[name = tensor<string, []>("cast_87")];
            tensor<fp16, [1, 32, 1, 1]> add_12_cast_fp16 = add(x = reduce_mean_20_to_fp16, y = add_12_y_0_to_fp16)[name = tensor<string, []>("add_12_cast_fp16")];
            tensor<fp16, [1, 32, 1, 1]> sqrt_6_cast_fp16 = sqrt(x = add_12_cast_fp16)[name = tensor<string, []>("sqrt_6_cast_fp16")];
            tensor<fp16, [1, 32, 32, ?]> real_div_6_cast_fp16 = real_div(x = sub_12_cast_fp16, y = sqrt_6_cast_fp16)[name = tensor<string, []>("real_div_6_cast_fp16")];
            tensor<fp16, [1, 1024, ?]> reshape_25_cast_fp16 = reshape(shape = shape_42_cast_fp16, x = real_div_6_cast_fp16)[name = tensor<string, []>("reshape_25_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> reshape_26_to_fp16 = const()[name = tensor<string, []>("reshape_26_to_fp16"), val = tensor<fp16, [1, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(359822272)))];
            tensor<fp16, [1, 1024, ?]> mul_42_cast_fp16 = mul(x = reshape_25_cast_fp16, y = reshape_26_to_fp16)[name = tensor<string, []>("mul_42_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> reshape_27_to_fp16 = const()[name = tensor<string, []>("reshape_27_to_fp16"), val = tensor<fp16, [1, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(359824384)))];
            tensor<fp16, [1, 1024, ?]> add_13_cast_fp16 = add(x = mul_42_cast_fp16, y = reshape_27_to_fp16)[name = tensor<string, []>("add_13_cast_fp16")];
            tensor<fp16, [1, 1024, ?]> input_161_cast_fp16 = silu(x = add_13_cast_fp16)[name = tensor<string, []>("input_161_cast_fp16")];
            tensor<string, []> input_163_pad_type_0 = const()[name = tensor<string, []>("input_163_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [2]> input_163_pad_0 = const()[name = tensor<string, []>("input_163_pad_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [1]> input_163_strides_0 = const()[name = tensor<string, []>("input_163_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> input_163_dilations_0 = const()[name = tensor<string, []>("input_163_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> input_163_groups_0 = const()[name = tensor<string, []>("input_163_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 3]> post_net_1_conv1_weight_to_fp16 = const()[name = tensor<string, []>("post_net_1_conv1_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 3]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(359826496)))];
            tensor<fp16, [1024]> post_net_1_conv1_bias_to_fp16 = const()[name = tensor<string, []>("post_net_1_conv1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(366118016)))];
            tensor<fp16, [1, 1024, ?]> input_163_cast_fp16 = conv(bias = post_net_1_conv1_bias_to_fp16, dilations = input_163_dilations_0, groups = input_163_groups_0, pad = input_163_pad_0, pad_type = input_163_pad_type_0, strides = input_163_strides_0, weight = post_net_1_conv1_weight_to_fp16, x = input_161_cast_fp16)[name = tensor<string, []>("input_163_cast_fp16")];
            tensor<int32, [3]> shape_43_cast_fp16 = shape(x = input_163_cast_fp16)[name = tensor<string, []>("shape_43_cast_fp16")];
            tensor<int32, [4]> concat_31x = const()[name = tensor<string, []>("concat_31x"), val = tensor<int32, [4]>([1, 32, 32, -1])];
            tensor<fp16, [1, 32, 32, ?]> reshape_28_cast_fp16 = reshape(shape = concat_31x, x = input_163_cast_fp16)[name = tensor<string, []>("reshape_28_cast_fp16")];
            tensor<string, []> reshape_28_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("reshape_28_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<int32, [2]> reduce_mean_21_axes_0 = const()[name = tensor<string, []>("reduce_mean_21_axes_0"), val = tensor<int32, [2]>([2, 3])];
            tensor<bool, []> reduce_mean_21_keep_dims_0 = const()[name = tensor<string, []>("reduce_mean_21_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 32, 32, ?]> reshape_28_cast_fp16_to_fp32 = cast(dtype = reshape_28_cast_fp16_to_fp32_dtype_0, x = reshape_28_cast_fp16)[name = tensor<string, []>("cast_86")];
            tensor<fp32, [1, 32, 1, 1]> reduce_mean_21 = reduce_mean(axes = reduce_mean_21_axes_0, keep_dims = reduce_mean_21_keep_dims_0, x = reshape_28_cast_fp16_to_fp32)[name = tensor<string, []>("reduce_mean_21")];
            tensor<string, []> reduce_mean_21_to_fp16_dtype_0 = const()[name = tensor<string, []>("reduce_mean_21_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 32, 1, 1]> reduce_mean_21_to_fp16 = cast(dtype = reduce_mean_21_to_fp16_dtype_0, x = reduce_mean_21)[name = tensor<string, []>("cast_85")];
            tensor<fp16, [1, 32, 32, ?]> sub_14_cast_fp16 = sub(x = reshape_28_cast_fp16, y = reduce_mean_21_to_fp16)[name = tensor<string, []>("sub_14_cast_fp16")];
            tensor<fp16, [1, 32, 32, ?]> square_7_cast_fp16 = square(x = sub_14_cast_fp16)[name = tensor<string, []>("square_7_cast_fp16")];
            tensor<string, []> square_7_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("square_7_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<int32, [2]> reduce_mean_23_axes_0 = const()[name = tensor<string, []>("reduce_mean_23_axes_0"), val = tensor<int32, [2]>([2, 3])];
            tensor<bool, []> reduce_mean_23_keep_dims_0 = const()[name = tensor<string, []>("reduce_mean_23_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 32, 32, ?]> square_7_cast_fp16_to_fp32 = cast(dtype = square_7_cast_fp16_to_fp32_dtype_0, x = square_7_cast_fp16)[name = tensor<string, []>("cast_84")];
            tensor<fp32, [1, 32, 1, 1]> reduce_mean_23 = reduce_mean(axes = reduce_mean_23_axes_0, keep_dims = reduce_mean_23_keep_dims_0, x = square_7_cast_fp16_to_fp32)[name = tensor<string, []>("reduce_mean_23")];
            tensor<string, []> reduce_mean_23_to_fp16_dtype_0 = const()[name = tensor<string, []>("reduce_mean_23_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> add_14_y_0_to_fp16 = const()[name = tensor<string, []>("add_14_y_0_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 32, 1, 1]> reduce_mean_23_to_fp16 = cast(dtype = reduce_mean_23_to_fp16_dtype_0, x = reduce_mean_23)[name = tensor<string, []>("cast_83")];
            tensor<fp16, [1, 32, 1, 1]> add_14_cast_fp16 = add(x = reduce_mean_23_to_fp16, y = add_14_y_0_to_fp16)[name = tensor<string, []>("add_14_cast_fp16")];
            tensor<fp16, [1, 32, 1, 1]> sqrt_7_cast_fp16 = sqrt(x = add_14_cast_fp16)[name = tensor<string, []>("sqrt_7_cast_fp16")];
            tensor<fp16, [1, 32, 32, ?]> real_div_7_cast_fp16 = real_div(x = sub_14_cast_fp16, y = sqrt_7_cast_fp16)[name = tensor<string, []>("real_div_7_cast_fp16")];
            tensor<fp16, [1, 1024, ?]> reshape_29_cast_fp16 = reshape(shape = shape_43_cast_fp16, x = real_div_7_cast_fp16)[name = tensor<string, []>("reshape_29_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> reshape_30_to_fp16 = const()[name = tensor<string, []>("reshape_30_to_fp16"), val = tensor<fp16, [1, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(366120128)))];
            tensor<fp16, [1, 1024, ?]> mul_43_cast_fp16 = mul(x = reshape_29_cast_fp16, y = reshape_30_to_fp16)[name = tensor<string, []>("mul_43_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> reshape_31_to_fp16 = const()[name = tensor<string, []>("reshape_31_to_fp16"), val = tensor<fp16, [1, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(366122240)))];
            tensor<fp16, [1, 1024, ?]> add_15_cast_fp16 = add(x = mul_43_cast_fp16, y = reshape_31_to_fp16)[name = tensor<string, []>("add_15_cast_fp16")];
            tensor<fp16, [1, 1024, ?]> input_165_cast_fp16 = silu(x = add_15_cast_fp16)[name = tensor<string, []>("input_165_cast_fp16")];
            tensor<string, []> h_pad_type_0 = const()[name = tensor<string, []>("h_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [2]> h_pad_0 = const()[name = tensor<string, []>("h_pad_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [1]> h_strides_0 = const()[name = tensor<string, []>("h_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> h_dilations_0 = const()[name = tensor<string, []>("h_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> h_groups_0 = const()[name = tensor<string, []>("h_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 1024, 3]> post_net_1_conv2_weight_to_fp16 = const()[name = tensor<string, []>("post_net_1_conv2_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 3]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(366124352)))];
            tensor<fp16, [1024]> post_net_1_conv2_bias_to_fp16 = const()[name = tensor<string, []>("post_net_1_conv2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(372415872)))];
            tensor<fp16, [1, 1024, ?]> h_cast_fp16 = conv(bias = post_net_1_conv2_bias_to_fp16, dilations = h_dilations_0, groups = h_groups_0, pad = h_pad_0, pad_type = h_pad_type_0, strides = h_strides_0, weight = post_net_1_conv2_weight_to_fp16, x = input_165_cast_fp16)[name = tensor<string, []>("h_cast_fp16")];
            tensor<fp16, [1, 1024, ?]> x_119_cast_fp16 = add(x = input_159_cast_fp16, y = h_cast_fp16)[name = tensor<string, []>("x_119_cast_fp16")];
            tensor<int32, [3]> input_169_perm_0 = const()[name = tensor<string, []>("input_169_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> input_169_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("input_169_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_1316 = const()[name = tensor<string, []>("op_1316"), val = tensor<fp32, []>(0x1.0c6f7ap-20)];
            tensor<int32, [1]> input_axes_0 = const()[name = tensor<string, []>("input_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, ?, 1024]> input_169_cast_fp16 = transpose(perm = input_169_perm_0, x = x_119_cast_fp16)[name = tensor<string, []>("transpose_1")];
            tensor<fp32, [1, ?, 1024]> input_169_cast_fp16_to_fp32 = cast(dtype = input_169_cast_fp16_to_fp32_dtype_0, x = input_169_cast_fp16)[name = tensor<string, []>("cast_82")];
            tensor<fp32, [1, ?, 1024]> input = layer_norm(axes = input_axes_0, beta = final_layer_norm_bias, epsilon = var_1316, gamma = final_layer_norm_weight, x = input_169_cast_fp16_to_fp32)[name = tensor<string, []>("input")];
            tensor<string, []> input_to_fp16_dtype_0 = const()[name = tensor<string, []>("input_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1922, 1024]> head_out_weight_to_fp16 = const()[name = tensor<string, []>("head_out_weight_to_fp16"), val = tensor<fp16, [1922, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(372417984)))];
            tensor<fp16, [1922]> head_out_bias_to_fp16 = const()[name = tensor<string, []>("head_out_bias_to_fp16"), val = tensor<fp16, [1922]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(376354304)))];
            tensor<fp16, [1, ?, 1024]> input_to_fp16 = cast(dtype = input_to_fp16_dtype_0, x = input)[name = tensor<string, []>("cast_81")];
            tensor<fp16, [1, ?, 1922]> linear_50_cast_fp16 = linear(bias = head_out_bias_to_fp16, weight = head_out_weight_to_fp16, x = input_to_fp16)[name = tensor<string, []>("linear_50_cast_fp16")];
            tensor<int32, [3]> x_perm_0 = const()[name = tensor<string, []>("x_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [2]> var_1328_split_sizes_0 = const()[name = tensor<string, []>("op_1328_split_sizes_0"), val = tensor<int32, [2]>([961, 961])];
            tensor<int32, []> var_1328_axis_0 = const()[name = tensor<string, []>("op_1328_axis_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1, 1922, ?]> x_cast_fp16 = transpose(perm = x_perm_0, x = linear_50_cast_fp16)[name = tensor<string, []>("transpose_0")];
            tensor<fp16, [1, 961, ?]> var_1328_cast_fp16_0, tensor<fp16, [1, 961, ?]> var_1328_cast_fp16_1 = split(axis = var_1328_axis_0, split_sizes = var_1328_split_sizes_0, x = x_cast_fp16)[name = tensor<string, []>("op_1328_cast_fp16")];
            tensor<string, []> var_1328_cast_fp16_0_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_1328_cast_fp16_0_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<string, []> var_1328_cast_fp16_1_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_1328_cast_fp16_1_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, [1, 961, ?]> var_1328_cast_fp16_0_to_fp32 = cast(dtype = var_1328_cast_fp16_0_to_fp32_dtype_0, x = var_1328_cast_fp16_0)[name = tensor<string, []>("cast_80")];
            tensor<fp32, [1, 961, ?]> var_1330 = exp(x = var_1328_cast_fp16_0_to_fp32)[name = tensor<string, []>("op_1330")];
            tensor<string, []> var_1330_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_1330_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> const_12_to_fp16 = const()[name = tensor<string, []>("const_12_to_fp16"), val = tensor<fp16, []>(-inf)];
            tensor<fp16, []> var_1332_to_fp16 = const()[name = tensor<string, []>("op_1332_to_fp16"), val = tensor<fp16, []>(0x1.9p+6)];
            tensor<fp16, [1, 961, ?]> var_1330_to_fp16 = cast(dtype = var_1330_to_fp16_dtype_0, x = var_1330)[name = tensor<string, []>("cast_78")];
            tensor<fp16, [1, 961, ?]> clip_0_cast_fp16 = clip(alpha = const_12_to_fp16, beta = var_1332_to_fp16, x = var_1330_to_fp16)[name = tensor<string, []>("clip_0_cast_fp16")];
            tensor<fp32, [1, 961, ?]> var_1328_cast_fp16_1_to_fp32 = cast(dtype = var_1328_cast_fp16_1_to_fp32_dtype_0, x = var_1328_cast_fp16_1)[name = tensor<string, []>("cast_79")];
            tensor<fp32, [1, 961, ?]> var_1334 = cos(x = var_1328_cast_fp16_1_to_fp32)[name = tensor<string, []>("op_1334")];
            tensor<string, []> var_1334_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_1334_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 961, ?]> var_1334_to_fp16 = cast(dtype = var_1334_to_fp16_dtype_0, x = var_1334)[name = tensor<string, []>("cast_77")];
            tensor<fp16, [1, 961, ?]> real_cast_fp16 = mul(x = clip_0_cast_fp16, y = var_1334_to_fp16)[name = tensor<string, []>("real_cast_fp16")];
            tensor<fp32, [1, 961, ?]> var_1336 = sin(x = var_1328_cast_fp16_1_to_fp32)[name = tensor<string, []>("op_1336")];
            tensor<string, []> var_1336_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_1336_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 961, ?]> var_1336_to_fp16 = cast(dtype = var_1336_to_fp16_dtype_0, x = var_1336)[name = tensor<string, []>("cast_76")];
            tensor<fp16, [1, 961, ?]> imag_cast_fp16 = mul(x = clip_0_cast_fp16, y = var_1336_to_fp16)[name = tensor<string, []>("imag_cast_fp16")];
            tensor<string, []> var_1356_pad_type_0 = const()[name = tensor<string, []>("op_1356_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [1]> var_1356_strides_0 = const()[name = tensor<string, []>("op_1356_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> var_1356_pad_0 = const()[name = tensor<string, []>("op_1356_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> var_1356_dilations_0 = const()[name = tensor<string, []>("op_1356_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> var_1356_groups_0 = const()[name = tensor<string, []>("op_1356_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1920, 961, 1]> istft_idft_real_to_fp16 = const()[name = tensor<string, []>("istft_idft_real_to_fp16"), val = tensor<fp16, [1920, 961, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(376358272)))];
            tensor<fp16, [1, 1920, ?]> var_1356_cast_fp16 = conv(dilations = var_1356_dilations_0, groups = var_1356_groups_0, pad = var_1356_pad_0, pad_type = var_1356_pad_type_0, strides = var_1356_strides_0, weight = istft_idft_real_to_fp16, x = real_cast_fp16)[name = tensor<string, []>("op_1356_cast_fp16")];
            tensor<string, []> var_1361_pad_type_0 = const()[name = tensor<string, []>("op_1361_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [1]> var_1361_strides_0 = const()[name = tensor<string, []>("op_1361_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> var_1361_pad_0 = const()[name = tensor<string, []>("op_1361_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> var_1361_dilations_0 = const()[name = tensor<string, []>("op_1361_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> var_1361_groups_0 = const()[name = tensor<string, []>("op_1361_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1920, 961, 1]> istft_idft_imag_to_fp16 = const()[name = tensor<string, []>("istft_idft_imag_to_fp16"), val = tensor<fp16, [1920, 961, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(380048576)))];
            tensor<fp16, [1, 1920, ?]> var_1361_cast_fp16 = conv(dilations = var_1361_dilations_0, groups = var_1361_groups_0, pad = var_1361_pad_0, pad_type = var_1361_pad_type_0, strides = var_1361_strides_0, weight = istft_idft_imag_to_fp16, x = imag_cast_fp16)[name = tensor<string, []>("op_1361_cast_fp16")];
            tensor<fp16, [1, 1920, ?]> frames_cast_fp16 = add(x = var_1356_cast_fp16, y = var_1361_cast_fp16)[name = tensor<string, []>("frames_cast_fp16")];
            tensor<string, []> audio_pad_type_0 = const()[name = tensor<string, []>("audio_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [1]> audio_strides_0 = const()[name = tensor<string, []>("audio_strides_0"), val = tensor<int32, [1]>([480])];
            tensor<int32, [2]> audio_pad_0 = const()[name = tensor<string, []>("audio_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> audio_dilations_0 = const()[name = tensor<string, []>("audio_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> audio_groups_0 = const()[name = tensor<string, []>("audio_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1920, 1, 1920]> istft_ola_kernel_to_fp16 = const()[name = tensor<string, []>("istft_ola_kernel_to_fp16"), val = tensor<fp16, [1920, 1, 1920]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(383738880)))];
            tensor<fp16, [1, 1, ?]> audio_cast_fp16 = conv_transpose(dilations = audio_dilations_0, groups = audio_groups_0, pad = audio_pad_0, pad_type = audio_pad_type_0, strides = audio_strides_0, weight = istft_ola_kernel_to_fp16, x = frames_cast_fp16)[name = tensor<string, []>("audio_cast_fp16")];
            tensor<int32, [3]> var_1369_begin_0 = const()[name = tensor<string, []>("op_1369_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_1369_end_0 = const()[name = tensor<string, []>("op_1369_end_0"), val = tensor<int32, [3]>([1, 1, 0])];
            tensor<bool, [3]> var_1369_end_mask_0 = const()[name = tensor<string, []>("op_1369_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 1, ?]> var_1369_cast_fp16 = slice_by_index(begin = var_1369_begin_0, end = var_1369_end_0, end_mask = var_1369_end_mask_0, x = real_cast_fp16)[name = tensor<string, []>("op_1369_cast_fp16")];
            tensor<fp16, []> fill_like_0_value_0_to_fp16 = const()[name = tensor<string, []>("fill_like_0_value_0_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [1, 1, ?]> fill_like_0_cast_fp16 = fill_like(ref_tensor = var_1369_cast_fp16, value = fill_like_0_value_0_to_fp16)[name = tensor<string, []>("fill_like_0_cast_fp16")];
            tensor<string, []> envelope_pad_type_0 = const()[name = tensor<string, []>("envelope_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [1]> envelope_strides_0 = const()[name = tensor<string, []>("envelope_strides_0"), val = tensor<int32, [1]>([480])];
            tensor<int32, [2]> envelope_pad_0 = const()[name = tensor<string, []>("envelope_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> envelope_dilations_0 = const()[name = tensor<string, []>("envelope_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, []> envelope_groups_0 = const()[name = tensor<string, []>("envelope_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1, 1, 1920]> istft_win_sq_to_fp16 = const()[name = tensor<string, []>("istft_win_sq_to_fp16"), val = tensor<fp16, [1, 1, 1920]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(391111744)))];
            tensor<fp16, [1, 1, ?]> envelope_cast_fp16 = conv_transpose(dilations = envelope_dilations_0, groups = envelope_groups_0, pad = envelope_pad_0, pad_type = envelope_pad_type_0, strides = envelope_strides_0, weight = istft_win_sq_to_fp16, x = fill_like_0_cast_fp16)[name = tensor<string, []>("envelope_cast_fp16")];
            tensor<int32, [3]> var_1378_begin_0 = const()[name = tensor<string, []>("op_1378_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_1378_end_0 = const()[name = tensor<string, []>("op_1378_end_0"), val = tensor<int32, [3]>([1, 1, 0])];
            tensor<bool, [3]> var_1378_end_mask_0 = const()[name = tensor<string, []>("op_1378_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<bool, [3]> var_1378_squeeze_mask_0 = const()[name = tensor<string, []>("op_1378_squeeze_mask_0"), val = tensor<bool, [3]>([false, true, false])];
            tensor<fp16, [1, ?]> var_1378_cast_fp16 = slice_by_index(begin = var_1378_begin_0, end = var_1378_end_0, end_mask = var_1378_end_mask_0, squeeze_mask = var_1378_squeeze_mask_0, x = audio_cast_fp16)[name = tensor<string, []>("op_1378_cast_fp16")];
            tensor<int32, [2]> var_1379_begin_0 = const()[name = tensor<string, []>("op_1379_begin_0"), val = tensor<int32, [2]>([0, 720])];
            tensor<int32, [2]> var_1379_end_0 = const()[name = tensor<string, []>("op_1379_end_0"), val = tensor<int32, [2]>([1, -720])];
            tensor<bool, [2]> var_1379_end_mask_0 = const()[name = tensor<string, []>("op_1379_end_mask_0"), val = tensor<bool, [2]>([true, false])];
            tensor<fp16, [1, ?]> var_1379_cast_fp16 = slice_by_index(begin = var_1379_begin_0, end = var_1379_end_0, end_mask = var_1379_end_mask_0, x = var_1378_cast_fp16)[name = tensor<string, []>("op_1379_cast_fp16")];
            tensor<int32, [3]> var_1381_begin_0 = const()[name = tensor<string, []>("op_1381_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_1381_end_0 = const()[name = tensor<string, []>("op_1381_end_0"), val = tensor<int32, [3]>([1, 1, 0])];
            tensor<bool, [3]> var_1381_end_mask_0 = const()[name = tensor<string, []>("op_1381_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<bool, [3]> var_1381_squeeze_mask_0 = const()[name = tensor<string, []>("op_1381_squeeze_mask_0"), val = tensor<bool, [3]>([false, true, false])];
            tensor<fp16, [1, ?]> var_1381_cast_fp16 = slice_by_index(begin = var_1381_begin_0, end = var_1381_end_0, end_mask = var_1381_end_mask_0, squeeze_mask = var_1381_squeeze_mask_0, x = envelope_cast_fp16)[name = tensor<string, []>("op_1381_cast_fp16")];
            tensor<int32, [2]> var_1382_begin_0 = const()[name = tensor<string, []>("op_1382_begin_0"), val = tensor<int32, [2]>([0, 720])];
            tensor<int32, [2]> var_1382_end_0 = const()[name = tensor<string, []>("op_1382_end_0"), val = tensor<int32, [2]>([1, -720])];
            tensor<bool, [2]> var_1382_end_mask_0 = const()[name = tensor<string, []>("op_1382_end_mask_0"), val = tensor<bool, [2]>([true, false])];
            tensor<fp16, [1, ?]> var_1382_cast_fp16 = slice_by_index(begin = var_1382_begin_0, end = var_1382_end_0, end_mask = var_1382_end_mask_0, x = var_1381_cast_fp16)[name = tensor<string, []>("op_1382_cast_fp16")];
            tensor<fp16, [1, ?]> var_1383_cast_fp16 = real_div(x = var_1379_cast_fp16, y = var_1382_cast_fp16)[name = tensor<string, []>("op_1383_cast_fp16")];
            tensor<string, []> var_1383_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_1383_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, [1, ?]> audio = cast(dtype = var_1383_cast_fp16_to_fp32_dtype_0, x = var_1383_cast_fp16)[name = tensor<string, []>("cast_75")];
        } -> (audio);
}